xref: /freebsd/sys/contrib/openzfs/module/zfs/vdev.c (revision 608da65de9552d5678c1000776ed69da04a45983)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2021 by Delphix. All rights reserved.
25  * Copyright 2017 Nexenta Systems, Inc.
26  * Copyright (c) 2014 Integros [integros.com]
27  * Copyright 2016 Toomas Soome <tsoome@me.com>
28  * Copyright 2017 Joyent, Inc.
29  * Copyright (c) 2017, Intel Corporation.
30  * Copyright (c) 2019, Datto Inc. All rights reserved.
31  * Copyright (c) 2021, Klara Inc.
32  * Copyright (c) 2021, 2023 Hewlett Packard Enterprise Development LP.
33  */
34 
35 #include <sys/zfs_context.h>
36 #include <sys/fm/fs/zfs.h>
37 #include <sys/spa.h>
38 #include <sys/spa_impl.h>
39 #include <sys/bpobj.h>
40 #include <sys/dmu.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_dir.h>
43 #include <sys/vdev_impl.h>
44 #include <sys/vdev_rebuild.h>
45 #include <sys/vdev_draid.h>
46 #include <sys/uberblock_impl.h>
47 #include <sys/metaslab.h>
48 #include <sys/metaslab_impl.h>
49 #include <sys/space_map.h>
50 #include <sys/space_reftree.h>
51 #include <sys/zio.h>
52 #include <sys/zap.h>
53 #include <sys/fs/zfs.h>
54 #include <sys/arc.h>
55 #include <sys/zil.h>
56 #include <sys/dsl_scan.h>
57 #include <sys/vdev_raidz.h>
58 #include <sys/abd.h>
59 #include <sys/vdev_initialize.h>
60 #include <sys/vdev_trim.h>
61 #include <sys/zvol.h>
62 #include <sys/zfs_ratelimit.h>
63 #include "zfs_prop.h"
64 
65 /*
66  * One metaslab from each (normal-class) vdev is used by the ZIL.  These are
67  * called "embedded slog metaslabs", are referenced by vdev_log_mg, and are
68  * part of the spa_embedded_log_class.  The metaslab with the most free space
69  * in each vdev is selected for this purpose when the pool is opened (or a
70  * vdev is added).  See vdev_metaslab_init().
71  *
72  * Log blocks can be allocated from the following locations.  Each one is tried
73  * in order until the allocation succeeds:
74  * 1. dedicated log vdevs, aka "slog" (spa_log_class)
75  * 2. embedded slog metaslabs (spa_embedded_log_class)
76  * 3. other metaslabs in normal vdevs (spa_normal_class)
77  *
78  * zfs_embedded_slog_min_ms disables the embedded slog if there are fewer
79  * than this number of metaslabs in the vdev.  This ensures that we don't set
80  * aside an unreasonable amount of space for the ZIL.  If set to less than
81  * 1 << (spa_slop_shift + 1), on small pools the usable space may be reduced
82  * (by more than 1<<spa_slop_shift) due to the embedded slog metaslab.
83  */
84 static uint_t zfs_embedded_slog_min_ms = 64;
85 
86 /* default target for number of metaslabs per top-level vdev */
87 static uint_t zfs_vdev_default_ms_count = 200;
88 
89 /* minimum number of metaslabs per top-level vdev */
90 static uint_t zfs_vdev_min_ms_count = 16;
91 
92 /* practical upper limit of total metaslabs per top-level vdev */
93 static uint_t zfs_vdev_ms_count_limit = 1ULL << 17;
94 
95 /* lower limit for metaslab size (512M) */
96 static uint_t zfs_vdev_default_ms_shift = 29;
97 
98 /* upper limit for metaslab size (16G) */
99 static uint_t zfs_vdev_max_ms_shift = 34;
100 
101 int vdev_validate_skip = B_FALSE;
102 
103 /*
104  * Since the DTL space map of a vdev is not expected to have a lot of
105  * entries, we default its block size to 4K.
106  */
107 int zfs_vdev_dtl_sm_blksz = (1 << 12);
108 
109 /*
110  * Rate limit slow IO (delay) events to this many per second.
111  */
112 static unsigned int zfs_slow_io_events_per_second = 20;
113 
114 /*
115  * Rate limit checksum events after this many checksum errors per second.
116  */
117 static unsigned int zfs_checksum_events_per_second = 20;
118 
119 /*
120  * Ignore errors during scrub/resilver.  Allows to work around resilver
121  * upon import when there are pool errors.
122  */
123 static int zfs_scan_ignore_errors = 0;
124 
125 /*
126  * vdev-wide space maps that have lots of entries written to them at
127  * the end of each transaction can benefit from a higher I/O bandwidth
128  * (e.g. vdev_obsolete_sm), thus we default their block size to 128K.
129  */
130 int zfs_vdev_standard_sm_blksz = (1 << 17);
131 
132 /*
133  * Tunable parameter for debugging or performance analysis. Setting this
134  * will cause pool corruption on power loss if a volatile out-of-order
135  * write cache is enabled.
136  */
137 int zfs_nocacheflush = 0;
138 
139 /*
140  * Maximum and minimum ashift values that can be automatically set based on
141  * vdev's physical ashift (disk's physical sector size).  While ASHIFT_MAX
142  * is higher than the maximum value, it is intentionally limited here to not
143  * excessively impact pool space efficiency.  Higher ashift values may still
144  * be forced by vdev logical ashift or by user via ashift property, but won't
145  * be set automatically as a performance optimization.
146  */
147 uint_t zfs_vdev_max_auto_ashift = 14;
148 uint_t zfs_vdev_min_auto_ashift = ASHIFT_MIN;
149 
150 void
151 vdev_dbgmsg(vdev_t *vd, const char *fmt, ...)
152 {
153 	va_list adx;
154 	char buf[256];
155 
156 	va_start(adx, fmt);
157 	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
158 	va_end(adx);
159 
160 	if (vd->vdev_path != NULL) {
161 		zfs_dbgmsg("%s vdev '%s': %s", vd->vdev_ops->vdev_op_type,
162 		    vd->vdev_path, buf);
163 	} else {
164 		zfs_dbgmsg("%s-%llu vdev (guid %llu): %s",
165 		    vd->vdev_ops->vdev_op_type,
166 		    (u_longlong_t)vd->vdev_id,
167 		    (u_longlong_t)vd->vdev_guid, buf);
168 	}
169 }
170 
171 void
172 vdev_dbgmsg_print_tree(vdev_t *vd, int indent)
173 {
174 	char state[20];
175 
176 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops) {
177 		zfs_dbgmsg("%*svdev %llu: %s", indent, "",
178 		    (u_longlong_t)vd->vdev_id,
179 		    vd->vdev_ops->vdev_op_type);
180 		return;
181 	}
182 
183 	switch (vd->vdev_state) {
184 	case VDEV_STATE_UNKNOWN:
185 		(void) snprintf(state, sizeof (state), "unknown");
186 		break;
187 	case VDEV_STATE_CLOSED:
188 		(void) snprintf(state, sizeof (state), "closed");
189 		break;
190 	case VDEV_STATE_OFFLINE:
191 		(void) snprintf(state, sizeof (state), "offline");
192 		break;
193 	case VDEV_STATE_REMOVED:
194 		(void) snprintf(state, sizeof (state), "removed");
195 		break;
196 	case VDEV_STATE_CANT_OPEN:
197 		(void) snprintf(state, sizeof (state), "can't open");
198 		break;
199 	case VDEV_STATE_FAULTED:
200 		(void) snprintf(state, sizeof (state), "faulted");
201 		break;
202 	case VDEV_STATE_DEGRADED:
203 		(void) snprintf(state, sizeof (state), "degraded");
204 		break;
205 	case VDEV_STATE_HEALTHY:
206 		(void) snprintf(state, sizeof (state), "healthy");
207 		break;
208 	default:
209 		(void) snprintf(state, sizeof (state), "<state %u>",
210 		    (uint_t)vd->vdev_state);
211 	}
212 
213 	zfs_dbgmsg("%*svdev %u: %s%s, guid: %llu, path: %s, %s", indent,
214 	    "", (int)vd->vdev_id, vd->vdev_ops->vdev_op_type,
215 	    vd->vdev_islog ? " (log)" : "",
216 	    (u_longlong_t)vd->vdev_guid,
217 	    vd->vdev_path ? vd->vdev_path : "N/A", state);
218 
219 	for (uint64_t i = 0; i < vd->vdev_children; i++)
220 		vdev_dbgmsg_print_tree(vd->vdev_child[i], indent + 2);
221 }
222 
223 /*
224  * Virtual device management.
225  */
226 
227 static vdev_ops_t *const vdev_ops_table[] = {
228 	&vdev_root_ops,
229 	&vdev_raidz_ops,
230 	&vdev_draid_ops,
231 	&vdev_draid_spare_ops,
232 	&vdev_mirror_ops,
233 	&vdev_replacing_ops,
234 	&vdev_spare_ops,
235 	&vdev_disk_ops,
236 	&vdev_file_ops,
237 	&vdev_missing_ops,
238 	&vdev_hole_ops,
239 	&vdev_indirect_ops,
240 	NULL
241 };
242 
243 /*
244  * Given a vdev type, return the appropriate ops vector.
245  */
246 static vdev_ops_t *
247 vdev_getops(const char *type)
248 {
249 	vdev_ops_t *ops, *const *opspp;
250 
251 	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
252 		if (strcmp(ops->vdev_op_type, type) == 0)
253 			break;
254 
255 	return (ops);
256 }
257 
258 /*
259  * Given a vdev and a metaslab class, find which metaslab group we're
260  * interested in. All vdevs may belong to two different metaslab classes.
261  * Dedicated slog devices use only the primary metaslab group, rather than a
262  * separate log group. For embedded slogs, the vdev_log_mg will be non-NULL.
263  */
264 metaslab_group_t *
265 vdev_get_mg(vdev_t *vd, metaslab_class_t *mc)
266 {
267 	if (mc == spa_embedded_log_class(vd->vdev_spa) &&
268 	    vd->vdev_log_mg != NULL)
269 		return (vd->vdev_log_mg);
270 	else
271 		return (vd->vdev_mg);
272 }
273 
274 void
275 vdev_default_xlate(vdev_t *vd, const range_seg64_t *logical_rs,
276     range_seg64_t *physical_rs, range_seg64_t *remain_rs)
277 {
278 	(void) vd, (void) remain_rs;
279 
280 	physical_rs->rs_start = logical_rs->rs_start;
281 	physical_rs->rs_end = logical_rs->rs_end;
282 }
283 
284 /*
285  * Derive the enumerated allocation bias from string input.
286  * String origin is either the per-vdev zap or zpool(8).
287  */
288 static vdev_alloc_bias_t
289 vdev_derive_alloc_bias(const char *bias)
290 {
291 	vdev_alloc_bias_t alloc_bias = VDEV_BIAS_NONE;
292 
293 	if (strcmp(bias, VDEV_ALLOC_BIAS_LOG) == 0)
294 		alloc_bias = VDEV_BIAS_LOG;
295 	else if (strcmp(bias, VDEV_ALLOC_BIAS_SPECIAL) == 0)
296 		alloc_bias = VDEV_BIAS_SPECIAL;
297 	else if (strcmp(bias, VDEV_ALLOC_BIAS_DEDUP) == 0)
298 		alloc_bias = VDEV_BIAS_DEDUP;
299 
300 	return (alloc_bias);
301 }
302 
303 /*
304  * Default asize function: return the MAX of psize with the asize of
305  * all children.  This is what's used by anything other than RAID-Z.
306  */
307 uint64_t
308 vdev_default_asize(vdev_t *vd, uint64_t psize)
309 {
310 	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
311 	uint64_t csize;
312 
313 	for (int c = 0; c < vd->vdev_children; c++) {
314 		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
315 		asize = MAX(asize, csize);
316 	}
317 
318 	return (asize);
319 }
320 
321 uint64_t
322 vdev_default_min_asize(vdev_t *vd)
323 {
324 	return (vd->vdev_min_asize);
325 }
326 
327 /*
328  * Get the minimum allocatable size. We define the allocatable size as
329  * the vdev's asize rounded to the nearest metaslab. This allows us to
330  * replace or attach devices which don't have the same physical size but
331  * can still satisfy the same number of allocations.
332  */
333 uint64_t
334 vdev_get_min_asize(vdev_t *vd)
335 {
336 	vdev_t *pvd = vd->vdev_parent;
337 
338 	/*
339 	 * If our parent is NULL (inactive spare or cache) or is the root,
340 	 * just return our own asize.
341 	 */
342 	if (pvd == NULL)
343 		return (vd->vdev_asize);
344 
345 	/*
346 	 * The top-level vdev just returns the allocatable size rounded
347 	 * to the nearest metaslab.
348 	 */
349 	if (vd == vd->vdev_top)
350 		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
351 
352 	return (pvd->vdev_ops->vdev_op_min_asize(pvd));
353 }
354 
355 void
356 vdev_set_min_asize(vdev_t *vd)
357 {
358 	vd->vdev_min_asize = vdev_get_min_asize(vd);
359 
360 	for (int c = 0; c < vd->vdev_children; c++)
361 		vdev_set_min_asize(vd->vdev_child[c]);
362 }
363 
364 /*
365  * Get the minimal allocation size for the top-level vdev.
366  */
367 uint64_t
368 vdev_get_min_alloc(vdev_t *vd)
369 {
370 	uint64_t min_alloc = 1ULL << vd->vdev_ashift;
371 
372 	if (vd->vdev_ops->vdev_op_min_alloc != NULL)
373 		min_alloc = vd->vdev_ops->vdev_op_min_alloc(vd);
374 
375 	return (min_alloc);
376 }
377 
378 /*
379  * Get the parity level for a top-level vdev.
380  */
381 uint64_t
382 vdev_get_nparity(vdev_t *vd)
383 {
384 	uint64_t nparity = 0;
385 
386 	if (vd->vdev_ops->vdev_op_nparity != NULL)
387 		nparity = vd->vdev_ops->vdev_op_nparity(vd);
388 
389 	return (nparity);
390 }
391 
392 static int
393 vdev_prop_get_int(vdev_t *vd, vdev_prop_t prop, uint64_t *value)
394 {
395 	spa_t *spa = vd->vdev_spa;
396 	objset_t *mos = spa->spa_meta_objset;
397 	uint64_t objid;
398 	int err;
399 
400 	if (vd->vdev_root_zap != 0) {
401 		objid = vd->vdev_root_zap;
402 	} else if (vd->vdev_top_zap != 0) {
403 		objid = vd->vdev_top_zap;
404 	} else if (vd->vdev_leaf_zap != 0) {
405 		objid = vd->vdev_leaf_zap;
406 	} else {
407 		return (EINVAL);
408 	}
409 
410 	err = zap_lookup(mos, objid, vdev_prop_to_name(prop),
411 	    sizeof (uint64_t), 1, value);
412 
413 	if (err == ENOENT)
414 		*value = vdev_prop_default_numeric(prop);
415 
416 	return (err);
417 }
418 
419 /*
420  * Get the number of data disks for a top-level vdev.
421  */
422 uint64_t
423 vdev_get_ndisks(vdev_t *vd)
424 {
425 	uint64_t ndisks = 1;
426 
427 	if (vd->vdev_ops->vdev_op_ndisks != NULL)
428 		ndisks = vd->vdev_ops->vdev_op_ndisks(vd);
429 
430 	return (ndisks);
431 }
432 
433 vdev_t *
434 vdev_lookup_top(spa_t *spa, uint64_t vdev)
435 {
436 	vdev_t *rvd = spa->spa_root_vdev;
437 
438 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
439 
440 	if (vdev < rvd->vdev_children) {
441 		ASSERT(rvd->vdev_child[vdev] != NULL);
442 		return (rvd->vdev_child[vdev]);
443 	}
444 
445 	return (NULL);
446 }
447 
448 vdev_t *
449 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
450 {
451 	vdev_t *mvd;
452 
453 	if (vd->vdev_guid == guid)
454 		return (vd);
455 
456 	for (int c = 0; c < vd->vdev_children; c++)
457 		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
458 		    NULL)
459 			return (mvd);
460 
461 	return (NULL);
462 }
463 
464 static int
465 vdev_count_leaves_impl(vdev_t *vd)
466 {
467 	int n = 0;
468 
469 	if (vd->vdev_ops->vdev_op_leaf)
470 		return (1);
471 
472 	for (int c = 0; c < vd->vdev_children; c++)
473 		n += vdev_count_leaves_impl(vd->vdev_child[c]);
474 
475 	return (n);
476 }
477 
478 int
479 vdev_count_leaves(spa_t *spa)
480 {
481 	int rc;
482 
483 	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
484 	rc = vdev_count_leaves_impl(spa->spa_root_vdev);
485 	spa_config_exit(spa, SCL_VDEV, FTAG);
486 
487 	return (rc);
488 }
489 
490 void
491 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
492 {
493 	size_t oldsize, newsize;
494 	uint64_t id = cvd->vdev_id;
495 	vdev_t **newchild;
496 
497 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
498 	ASSERT(cvd->vdev_parent == NULL);
499 
500 	cvd->vdev_parent = pvd;
501 
502 	if (pvd == NULL)
503 		return;
504 
505 	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
506 
507 	oldsize = pvd->vdev_children * sizeof (vdev_t *);
508 	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
509 	newsize = pvd->vdev_children * sizeof (vdev_t *);
510 
511 	newchild = kmem_alloc(newsize, KM_SLEEP);
512 	if (pvd->vdev_child != NULL) {
513 		memcpy(newchild, pvd->vdev_child, oldsize);
514 		kmem_free(pvd->vdev_child, oldsize);
515 	}
516 
517 	pvd->vdev_child = newchild;
518 	pvd->vdev_child[id] = cvd;
519 
520 	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
521 	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
522 
523 	/*
524 	 * Walk up all ancestors to update guid sum.
525 	 */
526 	for (; pvd != NULL; pvd = pvd->vdev_parent)
527 		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
528 
529 	if (cvd->vdev_ops->vdev_op_leaf) {
530 		list_insert_head(&cvd->vdev_spa->spa_leaf_list, cvd);
531 		cvd->vdev_spa->spa_leaf_list_gen++;
532 	}
533 }
534 
535 void
536 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
537 {
538 	int c;
539 	uint_t id = cvd->vdev_id;
540 
541 	ASSERT(cvd->vdev_parent == pvd);
542 
543 	if (pvd == NULL)
544 		return;
545 
546 	ASSERT(id < pvd->vdev_children);
547 	ASSERT(pvd->vdev_child[id] == cvd);
548 
549 	pvd->vdev_child[id] = NULL;
550 	cvd->vdev_parent = NULL;
551 
552 	for (c = 0; c < pvd->vdev_children; c++)
553 		if (pvd->vdev_child[c])
554 			break;
555 
556 	if (c == pvd->vdev_children) {
557 		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
558 		pvd->vdev_child = NULL;
559 		pvd->vdev_children = 0;
560 	}
561 
562 	if (cvd->vdev_ops->vdev_op_leaf) {
563 		spa_t *spa = cvd->vdev_spa;
564 		list_remove(&spa->spa_leaf_list, cvd);
565 		spa->spa_leaf_list_gen++;
566 	}
567 
568 	/*
569 	 * Walk up all ancestors to update guid sum.
570 	 */
571 	for (; pvd != NULL; pvd = pvd->vdev_parent)
572 		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
573 }
574 
575 /*
576  * Remove any holes in the child array.
577  */
578 void
579 vdev_compact_children(vdev_t *pvd)
580 {
581 	vdev_t **newchild, *cvd;
582 	int oldc = pvd->vdev_children;
583 	int newc;
584 
585 	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
586 
587 	if (oldc == 0)
588 		return;
589 
590 	for (int c = newc = 0; c < oldc; c++)
591 		if (pvd->vdev_child[c])
592 			newc++;
593 
594 	if (newc > 0) {
595 		newchild = kmem_zalloc(newc * sizeof (vdev_t *), KM_SLEEP);
596 
597 		for (int c = newc = 0; c < oldc; c++) {
598 			if ((cvd = pvd->vdev_child[c]) != NULL) {
599 				newchild[newc] = cvd;
600 				cvd->vdev_id = newc++;
601 			}
602 		}
603 	} else {
604 		newchild = NULL;
605 	}
606 
607 	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
608 	pvd->vdev_child = newchild;
609 	pvd->vdev_children = newc;
610 }
611 
612 /*
613  * Allocate and minimally initialize a vdev_t.
614  */
615 vdev_t *
616 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
617 {
618 	vdev_t *vd;
619 	vdev_indirect_config_t *vic;
620 
621 	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
622 	vic = &vd->vdev_indirect_config;
623 
624 	if (spa->spa_root_vdev == NULL) {
625 		ASSERT(ops == &vdev_root_ops);
626 		spa->spa_root_vdev = vd;
627 		spa->spa_load_guid = spa_generate_guid(NULL);
628 	}
629 
630 	if (guid == 0 && ops != &vdev_hole_ops) {
631 		if (spa->spa_root_vdev == vd) {
632 			/*
633 			 * The root vdev's guid will also be the pool guid,
634 			 * which must be unique among all pools.
635 			 */
636 			guid = spa_generate_guid(NULL);
637 		} else {
638 			/*
639 			 * Any other vdev's guid must be unique within the pool.
640 			 */
641 			guid = spa_generate_guid(spa);
642 		}
643 		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
644 	}
645 
646 	vd->vdev_spa = spa;
647 	vd->vdev_id = id;
648 	vd->vdev_guid = guid;
649 	vd->vdev_guid_sum = guid;
650 	vd->vdev_ops = ops;
651 	vd->vdev_state = VDEV_STATE_CLOSED;
652 	vd->vdev_ishole = (ops == &vdev_hole_ops);
653 	vic->vic_prev_indirect_vdev = UINT64_MAX;
654 
655 	rw_init(&vd->vdev_indirect_rwlock, NULL, RW_DEFAULT, NULL);
656 	mutex_init(&vd->vdev_obsolete_lock, NULL, MUTEX_DEFAULT, NULL);
657 	vd->vdev_obsolete_segments = range_tree_create(NULL, RANGE_SEG64, NULL,
658 	    0, 0);
659 
660 	/*
661 	 * Initialize rate limit structs for events.  We rate limit ZIO delay
662 	 * and checksum events so that we don't overwhelm ZED with thousands
663 	 * of events when a disk is acting up.
664 	 */
665 	zfs_ratelimit_init(&vd->vdev_delay_rl, &zfs_slow_io_events_per_second,
666 	    1);
667 	zfs_ratelimit_init(&vd->vdev_deadman_rl, &zfs_slow_io_events_per_second,
668 	    1);
669 	zfs_ratelimit_init(&vd->vdev_checksum_rl,
670 	    &zfs_checksum_events_per_second, 1);
671 
672 	/*
673 	 * Default Thresholds for tuning ZED
674 	 */
675 	vd->vdev_checksum_n = vdev_prop_default_numeric(VDEV_PROP_CHECKSUM_N);
676 	vd->vdev_checksum_t = vdev_prop_default_numeric(VDEV_PROP_CHECKSUM_T);
677 	vd->vdev_io_n = vdev_prop_default_numeric(VDEV_PROP_IO_N);
678 	vd->vdev_io_t = vdev_prop_default_numeric(VDEV_PROP_IO_T);
679 
680 	list_link_init(&vd->vdev_config_dirty_node);
681 	list_link_init(&vd->vdev_state_dirty_node);
682 	list_link_init(&vd->vdev_initialize_node);
683 	list_link_init(&vd->vdev_leaf_node);
684 	list_link_init(&vd->vdev_trim_node);
685 
686 	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_NOLOCKDEP, NULL);
687 	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
688 	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
689 	mutex_init(&vd->vdev_scan_io_queue_lock, NULL, MUTEX_DEFAULT, NULL);
690 
691 	mutex_init(&vd->vdev_initialize_lock, NULL, MUTEX_DEFAULT, NULL);
692 	mutex_init(&vd->vdev_initialize_io_lock, NULL, MUTEX_DEFAULT, NULL);
693 	cv_init(&vd->vdev_initialize_cv, NULL, CV_DEFAULT, NULL);
694 	cv_init(&vd->vdev_initialize_io_cv, NULL, CV_DEFAULT, NULL);
695 
696 	mutex_init(&vd->vdev_trim_lock, NULL, MUTEX_DEFAULT, NULL);
697 	mutex_init(&vd->vdev_autotrim_lock, NULL, MUTEX_DEFAULT, NULL);
698 	mutex_init(&vd->vdev_trim_io_lock, NULL, MUTEX_DEFAULT, NULL);
699 	cv_init(&vd->vdev_trim_cv, NULL, CV_DEFAULT, NULL);
700 	cv_init(&vd->vdev_autotrim_cv, NULL, CV_DEFAULT, NULL);
701 	cv_init(&vd->vdev_autotrim_kick_cv, NULL, CV_DEFAULT, NULL);
702 	cv_init(&vd->vdev_trim_io_cv, NULL, CV_DEFAULT, NULL);
703 
704 	mutex_init(&vd->vdev_rebuild_lock, NULL, MUTEX_DEFAULT, NULL);
705 	cv_init(&vd->vdev_rebuild_cv, NULL, CV_DEFAULT, NULL);
706 
707 	for (int t = 0; t < DTL_TYPES; t++) {
708 		vd->vdev_dtl[t] = range_tree_create(NULL, RANGE_SEG64, NULL, 0,
709 		    0);
710 	}
711 
712 	txg_list_create(&vd->vdev_ms_list, spa,
713 	    offsetof(struct metaslab, ms_txg_node));
714 	txg_list_create(&vd->vdev_dtl_list, spa,
715 	    offsetof(struct vdev, vdev_dtl_node));
716 	vd->vdev_stat.vs_timestamp = gethrtime();
717 	vdev_queue_init(vd);
718 
719 	return (vd);
720 }
721 
722 /*
723  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
724  * creating a new vdev or loading an existing one - the behavior is slightly
725  * different for each case.
726  */
727 int
728 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
729     int alloctype)
730 {
731 	vdev_ops_t *ops;
732 	const char *type;
733 	uint64_t guid = 0, islog;
734 	vdev_t *vd;
735 	vdev_indirect_config_t *vic;
736 	const char *tmp = NULL;
737 	int rc;
738 	vdev_alloc_bias_t alloc_bias = VDEV_BIAS_NONE;
739 	boolean_t top_level = (parent && !parent->vdev_parent);
740 
741 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
742 
743 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
744 		return (SET_ERROR(EINVAL));
745 
746 	if ((ops = vdev_getops(type)) == NULL)
747 		return (SET_ERROR(EINVAL));
748 
749 	/*
750 	 * If this is a load, get the vdev guid from the nvlist.
751 	 * Otherwise, vdev_alloc_common() will generate one for us.
752 	 */
753 	if (alloctype == VDEV_ALLOC_LOAD) {
754 		uint64_t label_id;
755 
756 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
757 		    label_id != id)
758 			return (SET_ERROR(EINVAL));
759 
760 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
761 			return (SET_ERROR(EINVAL));
762 	} else if (alloctype == VDEV_ALLOC_SPARE) {
763 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
764 			return (SET_ERROR(EINVAL));
765 	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
766 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
767 			return (SET_ERROR(EINVAL));
768 	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
769 		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
770 			return (SET_ERROR(EINVAL));
771 	}
772 
773 	/*
774 	 * The first allocated vdev must be of type 'root'.
775 	 */
776 	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
777 		return (SET_ERROR(EINVAL));
778 
779 	/*
780 	 * Determine whether we're a log vdev.
781 	 */
782 	islog = 0;
783 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
784 	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
785 		return (SET_ERROR(ENOTSUP));
786 
787 	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
788 		return (SET_ERROR(ENOTSUP));
789 
790 	if (top_level && alloctype == VDEV_ALLOC_ADD) {
791 		const char *bias;
792 
793 		/*
794 		 * If creating a top-level vdev, check for allocation
795 		 * classes input.
796 		 */
797 		if (nvlist_lookup_string(nv, ZPOOL_CONFIG_ALLOCATION_BIAS,
798 		    &bias) == 0) {
799 			alloc_bias = vdev_derive_alloc_bias(bias);
800 
801 			/* spa_vdev_add() expects feature to be enabled */
802 			if (spa->spa_load_state != SPA_LOAD_CREATE &&
803 			    !spa_feature_is_enabled(spa,
804 			    SPA_FEATURE_ALLOCATION_CLASSES)) {
805 				return (SET_ERROR(ENOTSUP));
806 			}
807 		}
808 
809 		/* spa_vdev_add() expects feature to be enabled */
810 		if (ops == &vdev_draid_ops &&
811 		    spa->spa_load_state != SPA_LOAD_CREATE &&
812 		    !spa_feature_is_enabled(spa, SPA_FEATURE_DRAID)) {
813 			return (SET_ERROR(ENOTSUP));
814 		}
815 	}
816 
817 	/*
818 	 * Initialize the vdev specific data.  This is done before calling
819 	 * vdev_alloc_common() since it may fail and this simplifies the
820 	 * error reporting and cleanup code paths.
821 	 */
822 	void *tsd = NULL;
823 	if (ops->vdev_op_init != NULL) {
824 		rc = ops->vdev_op_init(spa, nv, &tsd);
825 		if (rc != 0) {
826 			return (rc);
827 		}
828 	}
829 
830 	vd = vdev_alloc_common(spa, id, guid, ops);
831 	vd->vdev_tsd = tsd;
832 	vd->vdev_islog = islog;
833 
834 	if (top_level && alloc_bias != VDEV_BIAS_NONE)
835 		vd->vdev_alloc_bias = alloc_bias;
836 
837 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &tmp) == 0)
838 		vd->vdev_path = spa_strdup(tmp);
839 
840 	/*
841 	 * ZPOOL_CONFIG_AUX_STATE = "external" means we previously forced a
842 	 * fault on a vdev and want it to persist across imports (like with
843 	 * zpool offline -f).
844 	 */
845 	rc = nvlist_lookup_string(nv, ZPOOL_CONFIG_AUX_STATE, &tmp);
846 	if (rc == 0 && tmp != NULL && strcmp(tmp, "external") == 0) {
847 		vd->vdev_stat.vs_aux = VDEV_AUX_EXTERNAL;
848 		vd->vdev_faulted = 1;
849 		vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
850 	}
851 
852 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &tmp) == 0)
853 		vd->vdev_devid = spa_strdup(tmp);
854 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH, &tmp) == 0)
855 		vd->vdev_physpath = spa_strdup(tmp);
856 
857 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_VDEV_ENC_SYSFS_PATH,
858 	    &tmp) == 0)
859 		vd->vdev_enc_sysfs_path = spa_strdup(tmp);
860 
861 	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &tmp) == 0)
862 		vd->vdev_fru = spa_strdup(tmp);
863 
864 	/*
865 	 * Set the whole_disk property.  If it's not specified, leave the value
866 	 * as -1.
867 	 */
868 	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
869 	    &vd->vdev_wholedisk) != 0)
870 		vd->vdev_wholedisk = -1ULL;
871 
872 	vic = &vd->vdev_indirect_config;
873 
874 	ASSERT0(vic->vic_mapping_object);
875 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_OBJECT,
876 	    &vic->vic_mapping_object);
877 	ASSERT0(vic->vic_births_object);
878 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_BIRTHS,
879 	    &vic->vic_births_object);
880 	ASSERT3U(vic->vic_prev_indirect_vdev, ==, UINT64_MAX);
881 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_PREV_INDIRECT_VDEV,
882 	    &vic->vic_prev_indirect_vdev);
883 
884 	/*
885 	 * Look for the 'not present' flag.  This will only be set if the device
886 	 * was not present at the time of import.
887 	 */
888 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
889 	    &vd->vdev_not_present);
890 
891 	/*
892 	 * Get the alignment requirement. Ignore pool ashift for vdev
893 	 * attach case.
894 	 */
895 	if (alloctype != VDEV_ALLOC_ATTACH) {
896 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT,
897 		    &vd->vdev_ashift);
898 	} else {
899 		vd->vdev_attaching = B_TRUE;
900 	}
901 
902 	/*
903 	 * Retrieve the vdev creation time.
904 	 */
905 	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
906 	    &vd->vdev_crtxg);
907 
908 	if (vd->vdev_ops == &vdev_root_ops &&
909 	    (alloctype == VDEV_ALLOC_LOAD ||
910 	    alloctype == VDEV_ALLOC_SPLIT ||
911 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
912 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_ROOT_ZAP,
913 		    &vd->vdev_root_zap);
914 	}
915 
916 	/*
917 	 * If we're a top-level vdev, try to load the allocation parameters.
918 	 */
919 	if (top_level &&
920 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
921 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
922 		    &vd->vdev_ms_array);
923 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
924 		    &vd->vdev_ms_shift);
925 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
926 		    &vd->vdev_asize);
927 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NONALLOCATING,
928 		    &vd->vdev_noalloc);
929 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
930 		    &vd->vdev_removing);
931 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
932 		    &vd->vdev_top_zap);
933 	} else {
934 		ASSERT0(vd->vdev_top_zap);
935 	}
936 
937 	if (top_level && alloctype != VDEV_ALLOC_ATTACH) {
938 		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
939 		    alloctype == VDEV_ALLOC_ADD ||
940 		    alloctype == VDEV_ALLOC_SPLIT ||
941 		    alloctype == VDEV_ALLOC_ROOTPOOL);
942 		/* Note: metaslab_group_create() is now deferred */
943 	}
944 
945 	if (vd->vdev_ops->vdev_op_leaf &&
946 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
947 		(void) nvlist_lookup_uint64(nv,
948 		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
949 	} else {
950 		ASSERT0(vd->vdev_leaf_zap);
951 	}
952 
953 	/*
954 	 * If we're a leaf vdev, try to load the DTL object and other state.
955 	 */
956 
957 	if (vd->vdev_ops->vdev_op_leaf &&
958 	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
959 	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
960 		if (alloctype == VDEV_ALLOC_LOAD) {
961 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
962 			    &vd->vdev_dtl_object);
963 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
964 			    &vd->vdev_unspare);
965 		}
966 
967 		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
968 			uint64_t spare = 0;
969 
970 			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
971 			    &spare) == 0 && spare)
972 				spa_spare_add(vd);
973 		}
974 
975 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
976 		    &vd->vdev_offline);
977 
978 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
979 		    &vd->vdev_resilver_txg);
980 
981 		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REBUILD_TXG,
982 		    &vd->vdev_rebuild_txg);
983 
984 		if (nvlist_exists(nv, ZPOOL_CONFIG_RESILVER_DEFER))
985 			vdev_defer_resilver(vd);
986 
987 		/*
988 		 * In general, when importing a pool we want to ignore the
989 		 * persistent fault state, as the diagnosis made on another
990 		 * system may not be valid in the current context.  The only
991 		 * exception is if we forced a vdev to a persistently faulted
992 		 * state with 'zpool offline -f'.  The persistent fault will
993 		 * remain across imports until cleared.
994 		 *
995 		 * Local vdevs will remain in the faulted state.
996 		 */
997 		if (spa_load_state(spa) == SPA_LOAD_OPEN ||
998 		    spa_load_state(spa) == SPA_LOAD_IMPORT) {
999 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
1000 			    &vd->vdev_faulted);
1001 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
1002 			    &vd->vdev_degraded);
1003 			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
1004 			    &vd->vdev_removed);
1005 
1006 			if (vd->vdev_faulted || vd->vdev_degraded) {
1007 				const char *aux;
1008 
1009 				vd->vdev_label_aux =
1010 				    VDEV_AUX_ERR_EXCEEDED;
1011 				if (nvlist_lookup_string(nv,
1012 				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
1013 				    strcmp(aux, "external") == 0)
1014 					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
1015 				else
1016 					vd->vdev_faulted = 0ULL;
1017 			}
1018 		}
1019 	}
1020 
1021 	/*
1022 	 * Add ourselves to the parent's list of children.
1023 	 */
1024 	vdev_add_child(parent, vd);
1025 
1026 	*vdp = vd;
1027 
1028 	return (0);
1029 }
1030 
1031 void
1032 vdev_free(vdev_t *vd)
1033 {
1034 	spa_t *spa = vd->vdev_spa;
1035 
1036 	ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
1037 	ASSERT3P(vd->vdev_trim_thread, ==, NULL);
1038 	ASSERT3P(vd->vdev_autotrim_thread, ==, NULL);
1039 	ASSERT3P(vd->vdev_rebuild_thread, ==, NULL);
1040 
1041 	/*
1042 	 * Scan queues are normally destroyed at the end of a scan. If the
1043 	 * queue exists here, that implies the vdev is being removed while
1044 	 * the scan is still running.
1045 	 */
1046 	if (vd->vdev_scan_io_queue != NULL) {
1047 		mutex_enter(&vd->vdev_scan_io_queue_lock);
1048 		dsl_scan_io_queue_destroy(vd->vdev_scan_io_queue);
1049 		vd->vdev_scan_io_queue = NULL;
1050 		mutex_exit(&vd->vdev_scan_io_queue_lock);
1051 	}
1052 
1053 	/*
1054 	 * vdev_free() implies closing the vdev first.  This is simpler than
1055 	 * trying to ensure complicated semantics for all callers.
1056 	 */
1057 	vdev_close(vd);
1058 
1059 	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
1060 	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
1061 
1062 	/*
1063 	 * Free all children.
1064 	 */
1065 	for (int c = 0; c < vd->vdev_children; c++)
1066 		vdev_free(vd->vdev_child[c]);
1067 
1068 	ASSERT(vd->vdev_child == NULL);
1069 	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
1070 
1071 	if (vd->vdev_ops->vdev_op_fini != NULL)
1072 		vd->vdev_ops->vdev_op_fini(vd);
1073 
1074 	/*
1075 	 * Discard allocation state.
1076 	 */
1077 	if (vd->vdev_mg != NULL) {
1078 		vdev_metaslab_fini(vd);
1079 		metaslab_group_destroy(vd->vdev_mg);
1080 		vd->vdev_mg = NULL;
1081 	}
1082 	if (vd->vdev_log_mg != NULL) {
1083 		ASSERT0(vd->vdev_ms_count);
1084 		metaslab_group_destroy(vd->vdev_log_mg);
1085 		vd->vdev_log_mg = NULL;
1086 	}
1087 
1088 	ASSERT0(vd->vdev_stat.vs_space);
1089 	ASSERT0(vd->vdev_stat.vs_dspace);
1090 	ASSERT0(vd->vdev_stat.vs_alloc);
1091 
1092 	/*
1093 	 * Remove this vdev from its parent's child list.
1094 	 */
1095 	vdev_remove_child(vd->vdev_parent, vd);
1096 
1097 	ASSERT(vd->vdev_parent == NULL);
1098 	ASSERT(!list_link_active(&vd->vdev_leaf_node));
1099 
1100 	/*
1101 	 * Clean up vdev structure.
1102 	 */
1103 	vdev_queue_fini(vd);
1104 
1105 	if (vd->vdev_path)
1106 		spa_strfree(vd->vdev_path);
1107 	if (vd->vdev_devid)
1108 		spa_strfree(vd->vdev_devid);
1109 	if (vd->vdev_physpath)
1110 		spa_strfree(vd->vdev_physpath);
1111 
1112 	if (vd->vdev_enc_sysfs_path)
1113 		spa_strfree(vd->vdev_enc_sysfs_path);
1114 
1115 	if (vd->vdev_fru)
1116 		spa_strfree(vd->vdev_fru);
1117 
1118 	if (vd->vdev_isspare)
1119 		spa_spare_remove(vd);
1120 	if (vd->vdev_isl2cache)
1121 		spa_l2cache_remove(vd);
1122 
1123 	txg_list_destroy(&vd->vdev_ms_list);
1124 	txg_list_destroy(&vd->vdev_dtl_list);
1125 
1126 	mutex_enter(&vd->vdev_dtl_lock);
1127 	space_map_close(vd->vdev_dtl_sm);
1128 	for (int t = 0; t < DTL_TYPES; t++) {
1129 		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
1130 		range_tree_destroy(vd->vdev_dtl[t]);
1131 	}
1132 	mutex_exit(&vd->vdev_dtl_lock);
1133 
1134 	EQUIV(vd->vdev_indirect_births != NULL,
1135 	    vd->vdev_indirect_mapping != NULL);
1136 	if (vd->vdev_indirect_births != NULL) {
1137 		vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
1138 		vdev_indirect_births_close(vd->vdev_indirect_births);
1139 	}
1140 
1141 	if (vd->vdev_obsolete_sm != NULL) {
1142 		ASSERT(vd->vdev_removing ||
1143 		    vd->vdev_ops == &vdev_indirect_ops);
1144 		space_map_close(vd->vdev_obsolete_sm);
1145 		vd->vdev_obsolete_sm = NULL;
1146 	}
1147 	range_tree_destroy(vd->vdev_obsolete_segments);
1148 	rw_destroy(&vd->vdev_indirect_rwlock);
1149 	mutex_destroy(&vd->vdev_obsolete_lock);
1150 
1151 	mutex_destroy(&vd->vdev_dtl_lock);
1152 	mutex_destroy(&vd->vdev_stat_lock);
1153 	mutex_destroy(&vd->vdev_probe_lock);
1154 	mutex_destroy(&vd->vdev_scan_io_queue_lock);
1155 
1156 	mutex_destroy(&vd->vdev_initialize_lock);
1157 	mutex_destroy(&vd->vdev_initialize_io_lock);
1158 	cv_destroy(&vd->vdev_initialize_io_cv);
1159 	cv_destroy(&vd->vdev_initialize_cv);
1160 
1161 	mutex_destroy(&vd->vdev_trim_lock);
1162 	mutex_destroy(&vd->vdev_autotrim_lock);
1163 	mutex_destroy(&vd->vdev_trim_io_lock);
1164 	cv_destroy(&vd->vdev_trim_cv);
1165 	cv_destroy(&vd->vdev_autotrim_cv);
1166 	cv_destroy(&vd->vdev_autotrim_kick_cv);
1167 	cv_destroy(&vd->vdev_trim_io_cv);
1168 
1169 	mutex_destroy(&vd->vdev_rebuild_lock);
1170 	cv_destroy(&vd->vdev_rebuild_cv);
1171 
1172 	zfs_ratelimit_fini(&vd->vdev_delay_rl);
1173 	zfs_ratelimit_fini(&vd->vdev_deadman_rl);
1174 	zfs_ratelimit_fini(&vd->vdev_checksum_rl);
1175 
1176 	if (vd == spa->spa_root_vdev)
1177 		spa->spa_root_vdev = NULL;
1178 
1179 	kmem_free(vd, sizeof (vdev_t));
1180 }
1181 
1182 /*
1183  * Transfer top-level vdev state from svd to tvd.
1184  */
1185 static void
1186 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
1187 {
1188 	spa_t *spa = svd->vdev_spa;
1189 	metaslab_t *msp;
1190 	vdev_t *vd;
1191 	int t;
1192 
1193 	ASSERT(tvd == tvd->vdev_top);
1194 
1195 	tvd->vdev_ms_array = svd->vdev_ms_array;
1196 	tvd->vdev_ms_shift = svd->vdev_ms_shift;
1197 	tvd->vdev_ms_count = svd->vdev_ms_count;
1198 	tvd->vdev_top_zap = svd->vdev_top_zap;
1199 
1200 	svd->vdev_ms_array = 0;
1201 	svd->vdev_ms_shift = 0;
1202 	svd->vdev_ms_count = 0;
1203 	svd->vdev_top_zap = 0;
1204 
1205 	if (tvd->vdev_mg)
1206 		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
1207 	if (tvd->vdev_log_mg)
1208 		ASSERT3P(tvd->vdev_log_mg, ==, svd->vdev_log_mg);
1209 	tvd->vdev_mg = svd->vdev_mg;
1210 	tvd->vdev_log_mg = svd->vdev_log_mg;
1211 	tvd->vdev_ms = svd->vdev_ms;
1212 
1213 	svd->vdev_mg = NULL;
1214 	svd->vdev_log_mg = NULL;
1215 	svd->vdev_ms = NULL;
1216 
1217 	if (tvd->vdev_mg != NULL)
1218 		tvd->vdev_mg->mg_vd = tvd;
1219 	if (tvd->vdev_log_mg != NULL)
1220 		tvd->vdev_log_mg->mg_vd = tvd;
1221 
1222 	tvd->vdev_checkpoint_sm = svd->vdev_checkpoint_sm;
1223 	svd->vdev_checkpoint_sm = NULL;
1224 
1225 	tvd->vdev_alloc_bias = svd->vdev_alloc_bias;
1226 	svd->vdev_alloc_bias = VDEV_BIAS_NONE;
1227 
1228 	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
1229 	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
1230 	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
1231 
1232 	svd->vdev_stat.vs_alloc = 0;
1233 	svd->vdev_stat.vs_space = 0;
1234 	svd->vdev_stat.vs_dspace = 0;
1235 
1236 	/*
1237 	 * State which may be set on a top-level vdev that's in the
1238 	 * process of being removed.
1239 	 */
1240 	ASSERT0(tvd->vdev_indirect_config.vic_births_object);
1241 	ASSERT0(tvd->vdev_indirect_config.vic_mapping_object);
1242 	ASSERT3U(tvd->vdev_indirect_config.vic_prev_indirect_vdev, ==, -1ULL);
1243 	ASSERT3P(tvd->vdev_indirect_mapping, ==, NULL);
1244 	ASSERT3P(tvd->vdev_indirect_births, ==, NULL);
1245 	ASSERT3P(tvd->vdev_obsolete_sm, ==, NULL);
1246 	ASSERT0(tvd->vdev_noalloc);
1247 	ASSERT0(tvd->vdev_removing);
1248 	ASSERT0(tvd->vdev_rebuilding);
1249 	tvd->vdev_noalloc = svd->vdev_noalloc;
1250 	tvd->vdev_removing = svd->vdev_removing;
1251 	tvd->vdev_rebuilding = svd->vdev_rebuilding;
1252 	tvd->vdev_rebuild_config = svd->vdev_rebuild_config;
1253 	tvd->vdev_indirect_config = svd->vdev_indirect_config;
1254 	tvd->vdev_indirect_mapping = svd->vdev_indirect_mapping;
1255 	tvd->vdev_indirect_births = svd->vdev_indirect_births;
1256 	range_tree_swap(&svd->vdev_obsolete_segments,
1257 	    &tvd->vdev_obsolete_segments);
1258 	tvd->vdev_obsolete_sm = svd->vdev_obsolete_sm;
1259 	svd->vdev_indirect_config.vic_mapping_object = 0;
1260 	svd->vdev_indirect_config.vic_births_object = 0;
1261 	svd->vdev_indirect_config.vic_prev_indirect_vdev = -1ULL;
1262 	svd->vdev_indirect_mapping = NULL;
1263 	svd->vdev_indirect_births = NULL;
1264 	svd->vdev_obsolete_sm = NULL;
1265 	svd->vdev_noalloc = 0;
1266 	svd->vdev_removing = 0;
1267 	svd->vdev_rebuilding = 0;
1268 
1269 	for (t = 0; t < TXG_SIZE; t++) {
1270 		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
1271 			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
1272 		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
1273 			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
1274 		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
1275 			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
1276 	}
1277 
1278 	if (list_link_active(&svd->vdev_config_dirty_node)) {
1279 		vdev_config_clean(svd);
1280 		vdev_config_dirty(tvd);
1281 	}
1282 
1283 	if (list_link_active(&svd->vdev_state_dirty_node)) {
1284 		vdev_state_clean(svd);
1285 		vdev_state_dirty(tvd);
1286 	}
1287 
1288 	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
1289 	svd->vdev_deflate_ratio = 0;
1290 
1291 	tvd->vdev_islog = svd->vdev_islog;
1292 	svd->vdev_islog = 0;
1293 
1294 	dsl_scan_io_queue_vdev_xfer(svd, tvd);
1295 }
1296 
1297 static void
1298 vdev_top_update(vdev_t *tvd, vdev_t *vd)
1299 {
1300 	if (vd == NULL)
1301 		return;
1302 
1303 	vd->vdev_top = tvd;
1304 
1305 	for (int c = 0; c < vd->vdev_children; c++)
1306 		vdev_top_update(tvd, vd->vdev_child[c]);
1307 }
1308 
1309 /*
1310  * Add a mirror/replacing vdev above an existing vdev.  There is no need to
1311  * call .vdev_op_init() since mirror/replacing vdevs do not have private state.
1312  */
1313 vdev_t *
1314 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
1315 {
1316 	spa_t *spa = cvd->vdev_spa;
1317 	vdev_t *pvd = cvd->vdev_parent;
1318 	vdev_t *mvd;
1319 
1320 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1321 
1322 	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
1323 
1324 	mvd->vdev_asize = cvd->vdev_asize;
1325 	mvd->vdev_min_asize = cvd->vdev_min_asize;
1326 	mvd->vdev_max_asize = cvd->vdev_max_asize;
1327 	mvd->vdev_psize = cvd->vdev_psize;
1328 	mvd->vdev_ashift = cvd->vdev_ashift;
1329 	mvd->vdev_logical_ashift = cvd->vdev_logical_ashift;
1330 	mvd->vdev_physical_ashift = cvd->vdev_physical_ashift;
1331 	mvd->vdev_state = cvd->vdev_state;
1332 	mvd->vdev_crtxg = cvd->vdev_crtxg;
1333 
1334 	vdev_remove_child(pvd, cvd);
1335 	vdev_add_child(pvd, mvd);
1336 	cvd->vdev_id = mvd->vdev_children;
1337 	vdev_add_child(mvd, cvd);
1338 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
1339 
1340 	if (mvd == mvd->vdev_top)
1341 		vdev_top_transfer(cvd, mvd);
1342 
1343 	return (mvd);
1344 }
1345 
1346 /*
1347  * Remove a 1-way mirror/replacing vdev from the tree.
1348  */
1349 void
1350 vdev_remove_parent(vdev_t *cvd)
1351 {
1352 	vdev_t *mvd = cvd->vdev_parent;
1353 	vdev_t *pvd = mvd->vdev_parent;
1354 
1355 	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1356 
1357 	ASSERT(mvd->vdev_children == 1);
1358 	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
1359 	    mvd->vdev_ops == &vdev_replacing_ops ||
1360 	    mvd->vdev_ops == &vdev_spare_ops);
1361 	cvd->vdev_ashift = mvd->vdev_ashift;
1362 	cvd->vdev_logical_ashift = mvd->vdev_logical_ashift;
1363 	cvd->vdev_physical_ashift = mvd->vdev_physical_ashift;
1364 	vdev_remove_child(mvd, cvd);
1365 	vdev_remove_child(pvd, mvd);
1366 
1367 	/*
1368 	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
1369 	 * Otherwise, we could have detached an offline device, and when we
1370 	 * go to import the pool we'll think we have two top-level vdevs,
1371 	 * instead of a different version of the same top-level vdev.
1372 	 */
1373 	if (mvd->vdev_top == mvd) {
1374 		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
1375 		cvd->vdev_orig_guid = cvd->vdev_guid;
1376 		cvd->vdev_guid += guid_delta;
1377 		cvd->vdev_guid_sum += guid_delta;
1378 
1379 		/*
1380 		 * If pool not set for autoexpand, we need to also preserve
1381 		 * mvd's asize to prevent automatic expansion of cvd.
1382 		 * Otherwise if we are adjusting the mirror by attaching and
1383 		 * detaching children of non-uniform sizes, the mirror could
1384 		 * autoexpand, unexpectedly requiring larger devices to
1385 		 * re-establish the mirror.
1386 		 */
1387 		if (!cvd->vdev_spa->spa_autoexpand)
1388 			cvd->vdev_asize = mvd->vdev_asize;
1389 	}
1390 	cvd->vdev_id = mvd->vdev_id;
1391 	vdev_add_child(pvd, cvd);
1392 	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
1393 
1394 	if (cvd == cvd->vdev_top)
1395 		vdev_top_transfer(mvd, cvd);
1396 
1397 	ASSERT(mvd->vdev_children == 0);
1398 	vdev_free(mvd);
1399 }
1400 
1401 /*
1402  * Choose GCD for spa_gcd_alloc.
1403  */
1404 static uint64_t
1405 vdev_gcd(uint64_t a, uint64_t b)
1406 {
1407 	while (b != 0) {
1408 		uint64_t t = b;
1409 		b = a % b;
1410 		a = t;
1411 	}
1412 	return (a);
1413 }
1414 
1415 /*
1416  * Set spa_min_alloc and spa_gcd_alloc.
1417  */
1418 static void
1419 vdev_spa_set_alloc(spa_t *spa, uint64_t min_alloc)
1420 {
1421 	if (min_alloc < spa->spa_min_alloc)
1422 		spa->spa_min_alloc = min_alloc;
1423 	if (spa->spa_gcd_alloc == INT_MAX) {
1424 		spa->spa_gcd_alloc = min_alloc;
1425 	} else {
1426 		spa->spa_gcd_alloc = vdev_gcd(min_alloc,
1427 		    spa->spa_gcd_alloc);
1428 	}
1429 }
1430 
1431 void
1432 vdev_metaslab_group_create(vdev_t *vd)
1433 {
1434 	spa_t *spa = vd->vdev_spa;
1435 
1436 	/*
1437 	 * metaslab_group_create was delayed until allocation bias was available
1438 	 */
1439 	if (vd->vdev_mg == NULL) {
1440 		metaslab_class_t *mc;
1441 
1442 		if (vd->vdev_islog && vd->vdev_alloc_bias == VDEV_BIAS_NONE)
1443 			vd->vdev_alloc_bias = VDEV_BIAS_LOG;
1444 
1445 		ASSERT3U(vd->vdev_islog, ==,
1446 		    (vd->vdev_alloc_bias == VDEV_BIAS_LOG));
1447 
1448 		switch (vd->vdev_alloc_bias) {
1449 		case VDEV_BIAS_LOG:
1450 			mc = spa_log_class(spa);
1451 			break;
1452 		case VDEV_BIAS_SPECIAL:
1453 			mc = spa_special_class(spa);
1454 			break;
1455 		case VDEV_BIAS_DEDUP:
1456 			mc = spa_dedup_class(spa);
1457 			break;
1458 		default:
1459 			mc = spa_normal_class(spa);
1460 		}
1461 
1462 		vd->vdev_mg = metaslab_group_create(mc, vd,
1463 		    spa->spa_alloc_count);
1464 
1465 		if (!vd->vdev_islog) {
1466 			vd->vdev_log_mg = metaslab_group_create(
1467 			    spa_embedded_log_class(spa), vd, 1);
1468 		}
1469 
1470 		/*
1471 		 * The spa ashift min/max only apply for the normal metaslab
1472 		 * class. Class destination is late binding so ashift boundary
1473 		 * setting had to wait until now.
1474 		 */
1475 		if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1476 		    mc == spa_normal_class(spa) && vd->vdev_aux == NULL) {
1477 			if (vd->vdev_ashift > spa->spa_max_ashift)
1478 				spa->spa_max_ashift = vd->vdev_ashift;
1479 			if (vd->vdev_ashift < spa->spa_min_ashift)
1480 				spa->spa_min_ashift = vd->vdev_ashift;
1481 
1482 			uint64_t min_alloc = vdev_get_min_alloc(vd);
1483 			vdev_spa_set_alloc(spa, min_alloc);
1484 		}
1485 	}
1486 }
1487 
1488 int
1489 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
1490 {
1491 	spa_t *spa = vd->vdev_spa;
1492 	uint64_t oldc = vd->vdev_ms_count;
1493 	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
1494 	metaslab_t **mspp;
1495 	int error;
1496 	boolean_t expanding = (oldc != 0);
1497 
1498 	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
1499 
1500 	/*
1501 	 * This vdev is not being allocated from yet or is a hole.
1502 	 */
1503 	if (vd->vdev_ms_shift == 0)
1504 		return (0);
1505 
1506 	ASSERT(!vd->vdev_ishole);
1507 
1508 	ASSERT(oldc <= newc);
1509 
1510 	mspp = vmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
1511 
1512 	if (expanding) {
1513 		memcpy(mspp, vd->vdev_ms, oldc * sizeof (*mspp));
1514 		vmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
1515 	}
1516 
1517 	vd->vdev_ms = mspp;
1518 	vd->vdev_ms_count = newc;
1519 
1520 	for (uint64_t m = oldc; m < newc; m++) {
1521 		uint64_t object = 0;
1522 		/*
1523 		 * vdev_ms_array may be 0 if we are creating the "fake"
1524 		 * metaslabs for an indirect vdev for zdb's leak detection.
1525 		 * See zdb_leak_init().
1526 		 */
1527 		if (txg == 0 && vd->vdev_ms_array != 0) {
1528 			error = dmu_read(spa->spa_meta_objset,
1529 			    vd->vdev_ms_array,
1530 			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
1531 			    DMU_READ_PREFETCH);
1532 			if (error != 0) {
1533 				vdev_dbgmsg(vd, "unable to read the metaslab "
1534 				    "array [error=%d]", error);
1535 				return (error);
1536 			}
1537 		}
1538 
1539 		error = metaslab_init(vd->vdev_mg, m, object, txg,
1540 		    &(vd->vdev_ms[m]));
1541 		if (error != 0) {
1542 			vdev_dbgmsg(vd, "metaslab_init failed [error=%d]",
1543 			    error);
1544 			return (error);
1545 		}
1546 	}
1547 
1548 	/*
1549 	 * Find the emptiest metaslab on the vdev and mark it for use for
1550 	 * embedded slog by moving it from the regular to the log metaslab
1551 	 * group.
1552 	 */
1553 	if (vd->vdev_mg->mg_class == spa_normal_class(spa) &&
1554 	    vd->vdev_ms_count > zfs_embedded_slog_min_ms &&
1555 	    avl_is_empty(&vd->vdev_log_mg->mg_metaslab_tree)) {
1556 		uint64_t slog_msid = 0;
1557 		uint64_t smallest = UINT64_MAX;
1558 
1559 		/*
1560 		 * Note, we only search the new metaslabs, because the old
1561 		 * (pre-existing) ones may be active (e.g. have non-empty
1562 		 * range_tree's), and we don't move them to the new
1563 		 * metaslab_t.
1564 		 */
1565 		for (uint64_t m = oldc; m < newc; m++) {
1566 			uint64_t alloc =
1567 			    space_map_allocated(vd->vdev_ms[m]->ms_sm);
1568 			if (alloc < smallest) {
1569 				slog_msid = m;
1570 				smallest = alloc;
1571 			}
1572 		}
1573 		metaslab_t *slog_ms = vd->vdev_ms[slog_msid];
1574 		/*
1575 		 * The metaslab was marked as dirty at the end of
1576 		 * metaslab_init(). Remove it from the dirty list so that we
1577 		 * can uninitialize and reinitialize it to the new class.
1578 		 */
1579 		if (txg != 0) {
1580 			(void) txg_list_remove_this(&vd->vdev_ms_list,
1581 			    slog_ms, txg);
1582 		}
1583 		uint64_t sm_obj = space_map_object(slog_ms->ms_sm);
1584 		metaslab_fini(slog_ms);
1585 		VERIFY0(metaslab_init(vd->vdev_log_mg, slog_msid, sm_obj, txg,
1586 		    &vd->vdev_ms[slog_msid]));
1587 	}
1588 
1589 	if (txg == 0)
1590 		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
1591 
1592 	/*
1593 	 * If the vdev is marked as non-allocating then don't
1594 	 * activate the metaslabs since we want to ensure that
1595 	 * no allocations are performed on this device.
1596 	 */
1597 	if (vd->vdev_noalloc) {
1598 		/* track non-allocating vdev space */
1599 		spa->spa_nonallocating_dspace += spa_deflate(spa) ?
1600 		    vd->vdev_stat.vs_dspace : vd->vdev_stat.vs_space;
1601 	} else if (!expanding) {
1602 		metaslab_group_activate(vd->vdev_mg);
1603 		if (vd->vdev_log_mg != NULL)
1604 			metaslab_group_activate(vd->vdev_log_mg);
1605 	}
1606 
1607 	if (txg == 0)
1608 		spa_config_exit(spa, SCL_ALLOC, FTAG);
1609 
1610 	return (0);
1611 }
1612 
1613 void
1614 vdev_metaslab_fini(vdev_t *vd)
1615 {
1616 	if (vd->vdev_checkpoint_sm != NULL) {
1617 		ASSERT(spa_feature_is_active(vd->vdev_spa,
1618 		    SPA_FEATURE_POOL_CHECKPOINT));
1619 		space_map_close(vd->vdev_checkpoint_sm);
1620 		/*
1621 		 * Even though we close the space map, we need to set its
1622 		 * pointer to NULL. The reason is that vdev_metaslab_fini()
1623 		 * may be called multiple times for certain operations
1624 		 * (i.e. when destroying a pool) so we need to ensure that
1625 		 * this clause never executes twice. This logic is similar
1626 		 * to the one used for the vdev_ms clause below.
1627 		 */
1628 		vd->vdev_checkpoint_sm = NULL;
1629 	}
1630 
1631 	if (vd->vdev_ms != NULL) {
1632 		metaslab_group_t *mg = vd->vdev_mg;
1633 
1634 		metaslab_group_passivate(mg);
1635 		if (vd->vdev_log_mg != NULL) {
1636 			ASSERT(!vd->vdev_islog);
1637 			metaslab_group_passivate(vd->vdev_log_mg);
1638 		}
1639 
1640 		uint64_t count = vd->vdev_ms_count;
1641 		for (uint64_t m = 0; m < count; m++) {
1642 			metaslab_t *msp = vd->vdev_ms[m];
1643 			if (msp != NULL)
1644 				metaslab_fini(msp);
1645 		}
1646 		vmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
1647 		vd->vdev_ms = NULL;
1648 		vd->vdev_ms_count = 0;
1649 
1650 		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) {
1651 			ASSERT0(mg->mg_histogram[i]);
1652 			if (vd->vdev_log_mg != NULL)
1653 				ASSERT0(vd->vdev_log_mg->mg_histogram[i]);
1654 		}
1655 	}
1656 	ASSERT0(vd->vdev_ms_count);
1657 }
1658 
1659 typedef struct vdev_probe_stats {
1660 	boolean_t	vps_readable;
1661 	boolean_t	vps_writeable;
1662 	int		vps_flags;
1663 } vdev_probe_stats_t;
1664 
1665 static void
1666 vdev_probe_done(zio_t *zio)
1667 {
1668 	spa_t *spa = zio->io_spa;
1669 	vdev_t *vd = zio->io_vd;
1670 	vdev_probe_stats_t *vps = zio->io_private;
1671 
1672 	ASSERT(vd->vdev_probe_zio != NULL);
1673 
1674 	if (zio->io_type == ZIO_TYPE_READ) {
1675 		if (zio->io_error == 0)
1676 			vps->vps_readable = 1;
1677 		if (zio->io_error == 0 && spa_writeable(spa)) {
1678 			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
1679 			    zio->io_offset, zio->io_size, zio->io_abd,
1680 			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1681 			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
1682 		} else {
1683 			abd_free(zio->io_abd);
1684 		}
1685 	} else if (zio->io_type == ZIO_TYPE_WRITE) {
1686 		if (zio->io_error == 0)
1687 			vps->vps_writeable = 1;
1688 		abd_free(zio->io_abd);
1689 	} else if (zio->io_type == ZIO_TYPE_NULL) {
1690 		zio_t *pio;
1691 		zio_link_t *zl;
1692 
1693 		vd->vdev_cant_read |= !vps->vps_readable;
1694 		vd->vdev_cant_write |= !vps->vps_writeable;
1695 
1696 		if (vdev_readable(vd) &&
1697 		    (vdev_writeable(vd) || !spa_writeable(spa))) {
1698 			zio->io_error = 0;
1699 		} else {
1700 			ASSERT(zio->io_error != 0);
1701 			vdev_dbgmsg(vd, "failed probe");
1702 			(void) zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
1703 			    spa, vd, NULL, NULL, 0);
1704 			zio->io_error = SET_ERROR(ENXIO);
1705 		}
1706 
1707 		mutex_enter(&vd->vdev_probe_lock);
1708 		ASSERT(vd->vdev_probe_zio == zio);
1709 		vd->vdev_probe_zio = NULL;
1710 		mutex_exit(&vd->vdev_probe_lock);
1711 
1712 		zl = NULL;
1713 		while ((pio = zio_walk_parents(zio, &zl)) != NULL)
1714 			if (!vdev_accessible(vd, pio))
1715 				pio->io_error = SET_ERROR(ENXIO);
1716 
1717 		kmem_free(vps, sizeof (*vps));
1718 	}
1719 }
1720 
1721 /*
1722  * Determine whether this device is accessible.
1723  *
1724  * Read and write to several known locations: the pad regions of each
1725  * vdev label but the first, which we leave alone in case it contains
1726  * a VTOC.
1727  */
1728 zio_t *
1729 vdev_probe(vdev_t *vd, zio_t *zio)
1730 {
1731 	spa_t *spa = vd->vdev_spa;
1732 	vdev_probe_stats_t *vps = NULL;
1733 	zio_t *pio;
1734 
1735 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1736 
1737 	/*
1738 	 * Don't probe the probe.
1739 	 */
1740 	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1741 		return (NULL);
1742 
1743 	/*
1744 	 * To prevent 'probe storms' when a device fails, we create
1745 	 * just one probe i/o at a time.  All zios that want to probe
1746 	 * this vdev will become parents of the probe io.
1747 	 */
1748 	mutex_enter(&vd->vdev_probe_lock);
1749 
1750 	if ((pio = vd->vdev_probe_zio) == NULL) {
1751 		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1752 
1753 		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1754 		    ZIO_FLAG_DONT_AGGREGATE | ZIO_FLAG_TRYHARD;
1755 
1756 		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1757 			/*
1758 			 * vdev_cant_read and vdev_cant_write can only
1759 			 * transition from TRUE to FALSE when we have the
1760 			 * SCL_ZIO lock as writer; otherwise they can only
1761 			 * transition from FALSE to TRUE.  This ensures that
1762 			 * any zio looking at these values can assume that
1763 			 * failures persist for the life of the I/O.  That's
1764 			 * important because when a device has intermittent
1765 			 * connectivity problems, we want to ensure that
1766 			 * they're ascribed to the device (ENXIO) and not
1767 			 * the zio (EIO).
1768 			 *
1769 			 * Since we hold SCL_ZIO as writer here, clear both
1770 			 * values so the probe can reevaluate from first
1771 			 * principles.
1772 			 */
1773 			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1774 			vd->vdev_cant_read = B_FALSE;
1775 			vd->vdev_cant_write = B_FALSE;
1776 		}
1777 
1778 		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1779 		    vdev_probe_done, vps,
1780 		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1781 
1782 		/*
1783 		 * We can't change the vdev state in this context, so we
1784 		 * kick off an async task to do it on our behalf.
1785 		 */
1786 		if (zio != NULL) {
1787 			vd->vdev_probe_wanted = B_TRUE;
1788 			spa_async_request(spa, SPA_ASYNC_PROBE);
1789 		}
1790 	}
1791 
1792 	if (zio != NULL)
1793 		zio_add_child(zio, pio);
1794 
1795 	mutex_exit(&vd->vdev_probe_lock);
1796 
1797 	if (vps == NULL) {
1798 		ASSERT(zio != NULL);
1799 		return (NULL);
1800 	}
1801 
1802 	for (int l = 1; l < VDEV_LABELS; l++) {
1803 		zio_nowait(zio_read_phys(pio, vd,
1804 		    vdev_label_offset(vd->vdev_psize, l,
1805 		    offsetof(vdev_label_t, vl_be)), VDEV_PAD_SIZE,
1806 		    abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE),
1807 		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1808 		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1809 	}
1810 
1811 	if (zio == NULL)
1812 		return (pio);
1813 
1814 	zio_nowait(pio);
1815 	return (NULL);
1816 }
1817 
1818 static void
1819 vdev_load_child(void *arg)
1820 {
1821 	vdev_t *vd = arg;
1822 
1823 	vd->vdev_load_error = vdev_load(vd);
1824 }
1825 
1826 static void
1827 vdev_open_child(void *arg)
1828 {
1829 	vdev_t *vd = arg;
1830 
1831 	vd->vdev_open_thread = curthread;
1832 	vd->vdev_open_error = vdev_open(vd);
1833 	vd->vdev_open_thread = NULL;
1834 }
1835 
1836 static boolean_t
1837 vdev_uses_zvols(vdev_t *vd)
1838 {
1839 #ifdef _KERNEL
1840 	if (zvol_is_zvol(vd->vdev_path))
1841 		return (B_TRUE);
1842 #endif
1843 
1844 	for (int c = 0; c < vd->vdev_children; c++)
1845 		if (vdev_uses_zvols(vd->vdev_child[c]))
1846 			return (B_TRUE);
1847 
1848 	return (B_FALSE);
1849 }
1850 
1851 /*
1852  * Returns B_TRUE if the passed child should be opened.
1853  */
1854 static boolean_t
1855 vdev_default_open_children_func(vdev_t *vd)
1856 {
1857 	(void) vd;
1858 	return (B_TRUE);
1859 }
1860 
1861 /*
1862  * Open the requested child vdevs.  If any of the leaf vdevs are using
1863  * a ZFS volume then do the opens in a single thread.  This avoids a
1864  * deadlock when the current thread is holding the spa_namespace_lock.
1865  */
1866 static void
1867 vdev_open_children_impl(vdev_t *vd, vdev_open_children_func_t *open_func)
1868 {
1869 	int children = vd->vdev_children;
1870 
1871 	taskq_t *tq = taskq_create("vdev_open", children, minclsyspri,
1872 	    children, children, TASKQ_PREPOPULATE);
1873 	vd->vdev_nonrot = B_TRUE;
1874 
1875 	for (int c = 0; c < children; c++) {
1876 		vdev_t *cvd = vd->vdev_child[c];
1877 
1878 		if (open_func(cvd) == B_FALSE)
1879 			continue;
1880 
1881 		if (tq == NULL || vdev_uses_zvols(vd)) {
1882 			cvd->vdev_open_error = vdev_open(cvd);
1883 		} else {
1884 			VERIFY(taskq_dispatch(tq, vdev_open_child,
1885 			    cvd, TQ_SLEEP) != TASKQID_INVALID);
1886 		}
1887 
1888 		vd->vdev_nonrot &= cvd->vdev_nonrot;
1889 	}
1890 
1891 	if (tq != NULL) {
1892 		taskq_wait(tq);
1893 		taskq_destroy(tq);
1894 	}
1895 }
1896 
1897 /*
1898  * Open all child vdevs.
1899  */
1900 void
1901 vdev_open_children(vdev_t *vd)
1902 {
1903 	vdev_open_children_impl(vd, vdev_default_open_children_func);
1904 }
1905 
1906 /*
1907  * Conditionally open a subset of child vdevs.
1908  */
1909 void
1910 vdev_open_children_subset(vdev_t *vd, vdev_open_children_func_t *open_func)
1911 {
1912 	vdev_open_children_impl(vd, open_func);
1913 }
1914 
1915 /*
1916  * Compute the raidz-deflation ratio.  Note, we hard-code
1917  * in 128k (1 << 17) because it is the "typical" blocksize.
1918  * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
1919  * otherwise it would inconsistently account for existing bp's.
1920  */
1921 static void
1922 vdev_set_deflate_ratio(vdev_t *vd)
1923 {
1924 	if (vd == vd->vdev_top && !vd->vdev_ishole && vd->vdev_ashift != 0) {
1925 		vd->vdev_deflate_ratio = (1 << 17) /
1926 		    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
1927 	}
1928 }
1929 
1930 /*
1931  * Choose the best of two ashifts, preferring one between logical ashift
1932  * (absolute minimum) and administrator defined maximum, otherwise take
1933  * the biggest of the two.
1934  */
1935 uint64_t
1936 vdev_best_ashift(uint64_t logical, uint64_t a, uint64_t b)
1937 {
1938 	if (a > logical && a <= zfs_vdev_max_auto_ashift) {
1939 		if (b <= logical || b > zfs_vdev_max_auto_ashift)
1940 			return (a);
1941 		else
1942 			return (MAX(a, b));
1943 	} else if (b <= logical || b > zfs_vdev_max_auto_ashift)
1944 		return (MAX(a, b));
1945 	return (b);
1946 }
1947 
1948 /*
1949  * Maximize performance by inflating the configured ashift for top level
1950  * vdevs to be as close to the physical ashift as possible while maintaining
1951  * administrator defined limits and ensuring it doesn't go below the
1952  * logical ashift.
1953  */
1954 static void
1955 vdev_ashift_optimize(vdev_t *vd)
1956 {
1957 	ASSERT(vd == vd->vdev_top);
1958 
1959 	if (vd->vdev_ashift < vd->vdev_physical_ashift &&
1960 	    vd->vdev_physical_ashift <= zfs_vdev_max_auto_ashift) {
1961 		vd->vdev_ashift = MIN(
1962 		    MAX(zfs_vdev_max_auto_ashift, vd->vdev_ashift),
1963 		    MAX(zfs_vdev_min_auto_ashift,
1964 		    vd->vdev_physical_ashift));
1965 	} else {
1966 		/*
1967 		 * If the logical and physical ashifts are the same, then
1968 		 * we ensure that the top-level vdev's ashift is not smaller
1969 		 * than our minimum ashift value. For the unusual case
1970 		 * where logical ashift > physical ashift, we can't cap
1971 		 * the calculated ashift based on max ashift as that
1972 		 * would cause failures.
1973 		 * We still check if we need to increase it to match
1974 		 * the min ashift.
1975 		 */
1976 		vd->vdev_ashift = MAX(zfs_vdev_min_auto_ashift,
1977 		    vd->vdev_ashift);
1978 	}
1979 }
1980 
1981 /*
1982  * Prepare a virtual device for access.
1983  */
1984 int
1985 vdev_open(vdev_t *vd)
1986 {
1987 	spa_t *spa = vd->vdev_spa;
1988 	int error;
1989 	uint64_t osize = 0;
1990 	uint64_t max_osize = 0;
1991 	uint64_t asize, max_asize, psize;
1992 	uint64_t logical_ashift = 0;
1993 	uint64_t physical_ashift = 0;
1994 
1995 	ASSERT(vd->vdev_open_thread == curthread ||
1996 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1997 	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1998 	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1999 	    vd->vdev_state == VDEV_STATE_OFFLINE);
2000 
2001 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2002 	vd->vdev_cant_read = B_FALSE;
2003 	vd->vdev_cant_write = B_FALSE;
2004 	vd->vdev_min_asize = vdev_get_min_asize(vd);
2005 
2006 	/*
2007 	 * If this vdev is not removed, check its fault status.  If it's
2008 	 * faulted, bail out of the open.
2009 	 */
2010 	if (!vd->vdev_removed && vd->vdev_faulted) {
2011 		ASSERT(vd->vdev_children == 0);
2012 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
2013 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
2014 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
2015 		    vd->vdev_label_aux);
2016 		return (SET_ERROR(ENXIO));
2017 	} else if (vd->vdev_offline) {
2018 		ASSERT(vd->vdev_children == 0);
2019 		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
2020 		return (SET_ERROR(ENXIO));
2021 	}
2022 
2023 	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize,
2024 	    &logical_ashift, &physical_ashift);
2025 
2026 	/* Keep the device in removed state if unplugged */
2027 	if (error == ENOENT && vd->vdev_removed) {
2028 		vdev_set_state(vd, B_TRUE, VDEV_STATE_REMOVED,
2029 		    VDEV_AUX_NONE);
2030 		return (error);
2031 	}
2032 
2033 	/*
2034 	 * Physical volume size should never be larger than its max size, unless
2035 	 * the disk has shrunk while we were reading it or the device is buggy
2036 	 * or damaged: either way it's not safe for use, bail out of the open.
2037 	 */
2038 	if (osize > max_osize) {
2039 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2040 		    VDEV_AUX_OPEN_FAILED);
2041 		return (SET_ERROR(ENXIO));
2042 	}
2043 
2044 	/*
2045 	 * Reset the vdev_reopening flag so that we actually close
2046 	 * the vdev on error.
2047 	 */
2048 	vd->vdev_reopening = B_FALSE;
2049 	if (zio_injection_enabled && error == 0)
2050 		error = zio_handle_device_injection(vd, NULL, SET_ERROR(ENXIO));
2051 
2052 	if (error) {
2053 		if (vd->vdev_removed &&
2054 		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
2055 			vd->vdev_removed = B_FALSE;
2056 
2057 		if (vd->vdev_stat.vs_aux == VDEV_AUX_CHILDREN_OFFLINE) {
2058 			vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE,
2059 			    vd->vdev_stat.vs_aux);
2060 		} else {
2061 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2062 			    vd->vdev_stat.vs_aux);
2063 		}
2064 		return (error);
2065 	}
2066 
2067 	vd->vdev_removed = B_FALSE;
2068 
2069 	/*
2070 	 * Recheck the faulted flag now that we have confirmed that
2071 	 * the vdev is accessible.  If we're faulted, bail.
2072 	 */
2073 	if (vd->vdev_faulted) {
2074 		ASSERT(vd->vdev_children == 0);
2075 		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
2076 		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
2077 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
2078 		    vd->vdev_label_aux);
2079 		return (SET_ERROR(ENXIO));
2080 	}
2081 
2082 	if (vd->vdev_degraded) {
2083 		ASSERT(vd->vdev_children == 0);
2084 		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
2085 		    VDEV_AUX_ERR_EXCEEDED);
2086 	} else {
2087 		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
2088 	}
2089 
2090 	/*
2091 	 * For hole or missing vdevs we just return success.
2092 	 */
2093 	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
2094 		return (0);
2095 
2096 	for (int c = 0; c < vd->vdev_children; c++) {
2097 		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
2098 			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
2099 			    VDEV_AUX_NONE);
2100 			break;
2101 		}
2102 	}
2103 
2104 	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
2105 	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
2106 
2107 	if (vd->vdev_children == 0) {
2108 		if (osize < SPA_MINDEVSIZE) {
2109 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2110 			    VDEV_AUX_TOO_SMALL);
2111 			return (SET_ERROR(EOVERFLOW));
2112 		}
2113 		psize = osize;
2114 		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
2115 		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
2116 		    VDEV_LABEL_END_SIZE);
2117 	} else {
2118 		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
2119 		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
2120 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2121 			    VDEV_AUX_TOO_SMALL);
2122 			return (SET_ERROR(EOVERFLOW));
2123 		}
2124 		psize = 0;
2125 		asize = osize;
2126 		max_asize = max_osize;
2127 	}
2128 
2129 	/*
2130 	 * If the vdev was expanded, record this so that we can re-create the
2131 	 * uberblock rings in labels {2,3}, during the next sync.
2132 	 */
2133 	if ((psize > vd->vdev_psize) && (vd->vdev_psize != 0))
2134 		vd->vdev_copy_uberblocks = B_TRUE;
2135 
2136 	vd->vdev_psize = psize;
2137 
2138 	/*
2139 	 * Make sure the allocatable size hasn't shrunk too much.
2140 	 */
2141 	if (asize < vd->vdev_min_asize) {
2142 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2143 		    VDEV_AUX_BAD_LABEL);
2144 		return (SET_ERROR(EINVAL));
2145 	}
2146 
2147 	/*
2148 	 * We can always set the logical/physical ashift members since
2149 	 * their values are only used to calculate the vdev_ashift when
2150 	 * the device is first added to the config. These values should
2151 	 * not be used for anything else since they may change whenever
2152 	 * the device is reopened and we don't store them in the label.
2153 	 */
2154 	vd->vdev_physical_ashift =
2155 	    MAX(physical_ashift, vd->vdev_physical_ashift);
2156 	vd->vdev_logical_ashift = MAX(logical_ashift,
2157 	    vd->vdev_logical_ashift);
2158 
2159 	if (vd->vdev_asize == 0) {
2160 		/*
2161 		 * This is the first-ever open, so use the computed values.
2162 		 * For compatibility, a different ashift can be requested.
2163 		 */
2164 		vd->vdev_asize = asize;
2165 		vd->vdev_max_asize = max_asize;
2166 
2167 		/*
2168 		 * If the vdev_ashift was not overridden at creation time,
2169 		 * then set it the logical ashift and optimize the ashift.
2170 		 */
2171 		if (vd->vdev_ashift == 0) {
2172 			vd->vdev_ashift = vd->vdev_logical_ashift;
2173 
2174 			if (vd->vdev_logical_ashift > ASHIFT_MAX) {
2175 				vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2176 				    VDEV_AUX_ASHIFT_TOO_BIG);
2177 				return (SET_ERROR(EDOM));
2178 			}
2179 
2180 			if (vd->vdev_top == vd && vd->vdev_attaching == B_FALSE)
2181 				vdev_ashift_optimize(vd);
2182 			vd->vdev_attaching = B_FALSE;
2183 		}
2184 		if (vd->vdev_ashift != 0 && (vd->vdev_ashift < ASHIFT_MIN ||
2185 		    vd->vdev_ashift > ASHIFT_MAX)) {
2186 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2187 			    VDEV_AUX_BAD_ASHIFT);
2188 			return (SET_ERROR(EDOM));
2189 		}
2190 	} else {
2191 		/*
2192 		 * Make sure the alignment required hasn't increased.
2193 		 */
2194 		if (vd->vdev_ashift > vd->vdev_top->vdev_ashift &&
2195 		    vd->vdev_ops->vdev_op_leaf) {
2196 			(void) zfs_ereport_post(
2197 			    FM_EREPORT_ZFS_DEVICE_BAD_ASHIFT,
2198 			    spa, vd, NULL, NULL, 0);
2199 			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2200 			    VDEV_AUX_BAD_LABEL);
2201 			return (SET_ERROR(EDOM));
2202 		}
2203 		vd->vdev_max_asize = max_asize;
2204 	}
2205 
2206 	/*
2207 	 * If all children are healthy we update asize if either:
2208 	 * The asize has increased, due to a device expansion caused by dynamic
2209 	 * LUN growth or vdev replacement, and automatic expansion is enabled;
2210 	 * making the additional space available.
2211 	 *
2212 	 * The asize has decreased, due to a device shrink usually caused by a
2213 	 * vdev replace with a smaller device. This ensures that calculations
2214 	 * based of max_asize and asize e.g. esize are always valid. It's safe
2215 	 * to do this as we've already validated that asize is greater than
2216 	 * vdev_min_asize.
2217 	 */
2218 	if (vd->vdev_state == VDEV_STATE_HEALTHY &&
2219 	    ((asize > vd->vdev_asize &&
2220 	    (vd->vdev_expanding || spa->spa_autoexpand)) ||
2221 	    (asize < vd->vdev_asize)))
2222 		vd->vdev_asize = asize;
2223 
2224 	vdev_set_min_asize(vd);
2225 
2226 	/*
2227 	 * Ensure we can issue some IO before declaring the
2228 	 * vdev open for business.
2229 	 */
2230 	if (vd->vdev_ops->vdev_op_leaf &&
2231 	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
2232 		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
2233 		    VDEV_AUX_ERR_EXCEEDED);
2234 		return (error);
2235 	}
2236 
2237 	/*
2238 	 * Track the minimum allocation size.
2239 	 */
2240 	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
2241 	    vd->vdev_islog == 0 && vd->vdev_aux == NULL) {
2242 		uint64_t min_alloc = vdev_get_min_alloc(vd);
2243 		vdev_spa_set_alloc(spa, min_alloc);
2244 	}
2245 
2246 	/*
2247 	 * If this is a leaf vdev, assess whether a resilver is needed.
2248 	 * But don't do this if we are doing a reopen for a scrub, since
2249 	 * this would just restart the scrub we are already doing.
2250 	 */
2251 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen)
2252 		dsl_scan_assess_vdev(spa->spa_dsl_pool, vd);
2253 
2254 	return (0);
2255 }
2256 
2257 static void
2258 vdev_validate_child(void *arg)
2259 {
2260 	vdev_t *vd = arg;
2261 
2262 	vd->vdev_validate_thread = curthread;
2263 	vd->vdev_validate_error = vdev_validate(vd);
2264 	vd->vdev_validate_thread = NULL;
2265 }
2266 
2267 /*
2268  * Called once the vdevs are all opened, this routine validates the label
2269  * contents. This needs to be done before vdev_load() so that we don't
2270  * inadvertently do repair I/Os to the wrong device.
2271  *
2272  * This function will only return failure if one of the vdevs indicates that it
2273  * has since been destroyed or exported.  This is only possible if
2274  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
2275  * will be updated but the function will return 0.
2276  */
2277 int
2278 vdev_validate(vdev_t *vd)
2279 {
2280 	spa_t *spa = vd->vdev_spa;
2281 	taskq_t *tq = NULL;
2282 	nvlist_t *label;
2283 	uint64_t guid = 0, aux_guid = 0, top_guid;
2284 	uint64_t state;
2285 	nvlist_t *nvl;
2286 	uint64_t txg;
2287 	int children = vd->vdev_children;
2288 
2289 	if (vdev_validate_skip)
2290 		return (0);
2291 
2292 	if (children > 0) {
2293 		tq = taskq_create("vdev_validate", children, minclsyspri,
2294 		    children, children, TASKQ_PREPOPULATE);
2295 	}
2296 
2297 	for (uint64_t c = 0; c < children; c++) {
2298 		vdev_t *cvd = vd->vdev_child[c];
2299 
2300 		if (tq == NULL || vdev_uses_zvols(cvd)) {
2301 			vdev_validate_child(cvd);
2302 		} else {
2303 			VERIFY(taskq_dispatch(tq, vdev_validate_child, cvd,
2304 			    TQ_SLEEP) != TASKQID_INVALID);
2305 		}
2306 	}
2307 	if (tq != NULL) {
2308 		taskq_wait(tq);
2309 		taskq_destroy(tq);
2310 	}
2311 	for (int c = 0; c < children; c++) {
2312 		int error = vd->vdev_child[c]->vdev_validate_error;
2313 
2314 		if (error != 0)
2315 			return (SET_ERROR(EBADF));
2316 	}
2317 
2318 
2319 	/*
2320 	 * If the device has already failed, or was marked offline, don't do
2321 	 * any further validation.  Otherwise, label I/O will fail and we will
2322 	 * overwrite the previous state.
2323 	 */
2324 	if (!vd->vdev_ops->vdev_op_leaf || !vdev_readable(vd))
2325 		return (0);
2326 
2327 	/*
2328 	 * If we are performing an extreme rewind, we allow for a label that
2329 	 * was modified at a point after the current txg.
2330 	 * If config lock is not held do not check for the txg. spa_sync could
2331 	 * be updating the vdev's label before updating spa_last_synced_txg.
2332 	 */
2333 	if (spa->spa_extreme_rewind || spa_last_synced_txg(spa) == 0 ||
2334 	    spa_config_held(spa, SCL_CONFIG, RW_WRITER) != SCL_CONFIG)
2335 		txg = UINT64_MAX;
2336 	else
2337 		txg = spa_last_synced_txg(spa);
2338 
2339 	if ((label = vdev_label_read_config(vd, txg)) == NULL) {
2340 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2341 		    VDEV_AUX_BAD_LABEL);
2342 		vdev_dbgmsg(vd, "vdev_validate: failed reading config for "
2343 		    "txg %llu", (u_longlong_t)txg);
2344 		return (0);
2345 	}
2346 
2347 	/*
2348 	 * Determine if this vdev has been split off into another
2349 	 * pool.  If so, then refuse to open it.
2350 	 */
2351 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
2352 	    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
2353 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2354 		    VDEV_AUX_SPLIT_POOL);
2355 		nvlist_free(label);
2356 		vdev_dbgmsg(vd, "vdev_validate: vdev split into other pool");
2357 		return (0);
2358 	}
2359 
2360 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID, &guid) != 0) {
2361 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2362 		    VDEV_AUX_CORRUPT_DATA);
2363 		nvlist_free(label);
2364 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
2365 		    ZPOOL_CONFIG_POOL_GUID);
2366 		return (0);
2367 	}
2368 
2369 	/*
2370 	 * If config is not trusted then ignore the spa guid check. This is
2371 	 * necessary because if the machine crashed during a re-guid the new
2372 	 * guid might have been written to all of the vdev labels, but not the
2373 	 * cached config. The check will be performed again once we have the
2374 	 * trusted config from the MOS.
2375 	 */
2376 	if (spa->spa_trust_config && guid != spa_guid(spa)) {
2377 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2378 		    VDEV_AUX_CORRUPT_DATA);
2379 		nvlist_free(label);
2380 		vdev_dbgmsg(vd, "vdev_validate: vdev label pool_guid doesn't "
2381 		    "match config (%llu != %llu)", (u_longlong_t)guid,
2382 		    (u_longlong_t)spa_guid(spa));
2383 		return (0);
2384 	}
2385 
2386 	if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
2387 	    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
2388 	    &aux_guid) != 0)
2389 		aux_guid = 0;
2390 
2391 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0) {
2392 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2393 		    VDEV_AUX_CORRUPT_DATA);
2394 		nvlist_free(label);
2395 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
2396 		    ZPOOL_CONFIG_GUID);
2397 		return (0);
2398 	}
2399 
2400 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID, &top_guid)
2401 	    != 0) {
2402 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2403 		    VDEV_AUX_CORRUPT_DATA);
2404 		nvlist_free(label);
2405 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
2406 		    ZPOOL_CONFIG_TOP_GUID);
2407 		return (0);
2408 	}
2409 
2410 	/*
2411 	 * If this vdev just became a top-level vdev because its sibling was
2412 	 * detached, it will have adopted the parent's vdev guid -- but the
2413 	 * label may or may not be on disk yet. Fortunately, either version
2414 	 * of the label will have the same top guid, so if we're a top-level
2415 	 * vdev, we can safely compare to that instead.
2416 	 * However, if the config comes from a cachefile that failed to update
2417 	 * after the detach, a top-level vdev will appear as a non top-level
2418 	 * vdev in the config. Also relax the constraints if we perform an
2419 	 * extreme rewind.
2420 	 *
2421 	 * If we split this vdev off instead, then we also check the
2422 	 * original pool's guid. We don't want to consider the vdev
2423 	 * corrupt if it is partway through a split operation.
2424 	 */
2425 	if (vd->vdev_guid != guid && vd->vdev_guid != aux_guid) {
2426 		boolean_t mismatch = B_FALSE;
2427 		if (spa->spa_trust_config && !spa->spa_extreme_rewind) {
2428 			if (vd != vd->vdev_top || vd->vdev_guid != top_guid)
2429 				mismatch = B_TRUE;
2430 		} else {
2431 			if (vd->vdev_guid != top_guid &&
2432 			    vd->vdev_top->vdev_guid != guid)
2433 				mismatch = B_TRUE;
2434 		}
2435 
2436 		if (mismatch) {
2437 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2438 			    VDEV_AUX_CORRUPT_DATA);
2439 			nvlist_free(label);
2440 			vdev_dbgmsg(vd, "vdev_validate: config guid "
2441 			    "doesn't match label guid");
2442 			vdev_dbgmsg(vd, "CONFIG: guid %llu, top_guid %llu",
2443 			    (u_longlong_t)vd->vdev_guid,
2444 			    (u_longlong_t)vd->vdev_top->vdev_guid);
2445 			vdev_dbgmsg(vd, "LABEL: guid %llu, top_guid %llu, "
2446 			    "aux_guid %llu", (u_longlong_t)guid,
2447 			    (u_longlong_t)top_guid, (u_longlong_t)aux_guid);
2448 			return (0);
2449 		}
2450 	}
2451 
2452 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
2453 	    &state) != 0) {
2454 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2455 		    VDEV_AUX_CORRUPT_DATA);
2456 		nvlist_free(label);
2457 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
2458 		    ZPOOL_CONFIG_POOL_STATE);
2459 		return (0);
2460 	}
2461 
2462 	nvlist_free(label);
2463 
2464 	/*
2465 	 * If this is a verbatim import, no need to check the
2466 	 * state of the pool.
2467 	 */
2468 	if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
2469 	    spa_load_state(spa) == SPA_LOAD_OPEN &&
2470 	    state != POOL_STATE_ACTIVE) {
2471 		vdev_dbgmsg(vd, "vdev_validate: invalid pool state (%llu) "
2472 		    "for spa %s", (u_longlong_t)state, spa->spa_name);
2473 		return (SET_ERROR(EBADF));
2474 	}
2475 
2476 	/*
2477 	 * If we were able to open and validate a vdev that was
2478 	 * previously marked permanently unavailable, clear that state
2479 	 * now.
2480 	 */
2481 	if (vd->vdev_not_present)
2482 		vd->vdev_not_present = 0;
2483 
2484 	return (0);
2485 }
2486 
2487 static void
2488 vdev_copy_path_impl(vdev_t *svd, vdev_t *dvd)
2489 {
2490 	char *old, *new;
2491 	if (svd->vdev_path != NULL && dvd->vdev_path != NULL) {
2492 		if (strcmp(svd->vdev_path, dvd->vdev_path) != 0) {
2493 			zfs_dbgmsg("vdev_copy_path: vdev %llu: path changed "
2494 			    "from '%s' to '%s'", (u_longlong_t)dvd->vdev_guid,
2495 			    dvd->vdev_path, svd->vdev_path);
2496 			spa_strfree(dvd->vdev_path);
2497 			dvd->vdev_path = spa_strdup(svd->vdev_path);
2498 		}
2499 	} else if (svd->vdev_path != NULL) {
2500 		dvd->vdev_path = spa_strdup(svd->vdev_path);
2501 		zfs_dbgmsg("vdev_copy_path: vdev %llu: path set to '%s'",
2502 		    (u_longlong_t)dvd->vdev_guid, dvd->vdev_path);
2503 	}
2504 
2505 	/*
2506 	 * Our enclosure sysfs path may have changed between imports
2507 	 */
2508 	old = dvd->vdev_enc_sysfs_path;
2509 	new = svd->vdev_enc_sysfs_path;
2510 	if ((old != NULL && new == NULL) ||
2511 	    (old == NULL && new != NULL) ||
2512 	    ((old != NULL && new != NULL) && strcmp(new, old) != 0)) {
2513 		zfs_dbgmsg("vdev_copy_path: vdev %llu: vdev_enc_sysfs_path "
2514 		    "changed from '%s' to '%s'", (u_longlong_t)dvd->vdev_guid,
2515 		    old, new);
2516 
2517 		if (dvd->vdev_enc_sysfs_path)
2518 			spa_strfree(dvd->vdev_enc_sysfs_path);
2519 
2520 		if (svd->vdev_enc_sysfs_path) {
2521 			dvd->vdev_enc_sysfs_path = spa_strdup(
2522 			    svd->vdev_enc_sysfs_path);
2523 		} else {
2524 			dvd->vdev_enc_sysfs_path = NULL;
2525 		}
2526 	}
2527 }
2528 
2529 /*
2530  * Recursively copy vdev paths from one vdev to another. Source and destination
2531  * vdev trees must have same geometry otherwise return error. Intended to copy
2532  * paths from userland config into MOS config.
2533  */
2534 int
2535 vdev_copy_path_strict(vdev_t *svd, vdev_t *dvd)
2536 {
2537 	if ((svd->vdev_ops == &vdev_missing_ops) ||
2538 	    (svd->vdev_ishole && dvd->vdev_ishole) ||
2539 	    (dvd->vdev_ops == &vdev_indirect_ops))
2540 		return (0);
2541 
2542 	if (svd->vdev_ops != dvd->vdev_ops) {
2543 		vdev_dbgmsg(svd, "vdev_copy_path: vdev type mismatch: %s != %s",
2544 		    svd->vdev_ops->vdev_op_type, dvd->vdev_ops->vdev_op_type);
2545 		return (SET_ERROR(EINVAL));
2546 	}
2547 
2548 	if (svd->vdev_guid != dvd->vdev_guid) {
2549 		vdev_dbgmsg(svd, "vdev_copy_path: guids mismatch (%llu != "
2550 		    "%llu)", (u_longlong_t)svd->vdev_guid,
2551 		    (u_longlong_t)dvd->vdev_guid);
2552 		return (SET_ERROR(EINVAL));
2553 	}
2554 
2555 	if (svd->vdev_children != dvd->vdev_children) {
2556 		vdev_dbgmsg(svd, "vdev_copy_path: children count mismatch: "
2557 		    "%llu != %llu", (u_longlong_t)svd->vdev_children,
2558 		    (u_longlong_t)dvd->vdev_children);
2559 		return (SET_ERROR(EINVAL));
2560 	}
2561 
2562 	for (uint64_t i = 0; i < svd->vdev_children; i++) {
2563 		int error = vdev_copy_path_strict(svd->vdev_child[i],
2564 		    dvd->vdev_child[i]);
2565 		if (error != 0)
2566 			return (error);
2567 	}
2568 
2569 	if (svd->vdev_ops->vdev_op_leaf)
2570 		vdev_copy_path_impl(svd, dvd);
2571 
2572 	return (0);
2573 }
2574 
2575 static void
2576 vdev_copy_path_search(vdev_t *stvd, vdev_t *dvd)
2577 {
2578 	ASSERT(stvd->vdev_top == stvd);
2579 	ASSERT3U(stvd->vdev_id, ==, dvd->vdev_top->vdev_id);
2580 
2581 	for (uint64_t i = 0; i < dvd->vdev_children; i++) {
2582 		vdev_copy_path_search(stvd, dvd->vdev_child[i]);
2583 	}
2584 
2585 	if (!dvd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(dvd))
2586 		return;
2587 
2588 	/*
2589 	 * The idea here is that while a vdev can shift positions within
2590 	 * a top vdev (when replacing, attaching mirror, etc.) it cannot
2591 	 * step outside of it.
2592 	 */
2593 	vdev_t *vd = vdev_lookup_by_guid(stvd, dvd->vdev_guid);
2594 
2595 	if (vd == NULL || vd->vdev_ops != dvd->vdev_ops)
2596 		return;
2597 
2598 	ASSERT(vd->vdev_ops->vdev_op_leaf);
2599 
2600 	vdev_copy_path_impl(vd, dvd);
2601 }
2602 
2603 /*
2604  * Recursively copy vdev paths from one root vdev to another. Source and
2605  * destination vdev trees may differ in geometry. For each destination leaf
2606  * vdev, search a vdev with the same guid and top vdev id in the source.
2607  * Intended to copy paths from userland config into MOS config.
2608  */
2609 void
2610 vdev_copy_path_relaxed(vdev_t *srvd, vdev_t *drvd)
2611 {
2612 	uint64_t children = MIN(srvd->vdev_children, drvd->vdev_children);
2613 	ASSERT(srvd->vdev_ops == &vdev_root_ops);
2614 	ASSERT(drvd->vdev_ops == &vdev_root_ops);
2615 
2616 	for (uint64_t i = 0; i < children; i++) {
2617 		vdev_copy_path_search(srvd->vdev_child[i],
2618 		    drvd->vdev_child[i]);
2619 	}
2620 }
2621 
2622 /*
2623  * Close a virtual device.
2624  */
2625 void
2626 vdev_close(vdev_t *vd)
2627 {
2628 	vdev_t *pvd = vd->vdev_parent;
2629 	spa_t *spa __maybe_unused = vd->vdev_spa;
2630 
2631 	ASSERT(vd != NULL);
2632 	ASSERT(vd->vdev_open_thread == curthread ||
2633 	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2634 
2635 	/*
2636 	 * If our parent is reopening, then we are as well, unless we are
2637 	 * going offline.
2638 	 */
2639 	if (pvd != NULL && pvd->vdev_reopening)
2640 		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
2641 
2642 	vd->vdev_ops->vdev_op_close(vd);
2643 
2644 	/*
2645 	 * We record the previous state before we close it, so that if we are
2646 	 * doing a reopen(), we don't generate FMA ereports if we notice that
2647 	 * it's still faulted.
2648 	 */
2649 	vd->vdev_prevstate = vd->vdev_state;
2650 
2651 	if (vd->vdev_offline)
2652 		vd->vdev_state = VDEV_STATE_OFFLINE;
2653 	else
2654 		vd->vdev_state = VDEV_STATE_CLOSED;
2655 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
2656 }
2657 
2658 void
2659 vdev_hold(vdev_t *vd)
2660 {
2661 	spa_t *spa = vd->vdev_spa;
2662 
2663 	ASSERT(spa_is_root(spa));
2664 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
2665 		return;
2666 
2667 	for (int c = 0; c < vd->vdev_children; c++)
2668 		vdev_hold(vd->vdev_child[c]);
2669 
2670 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_ops->vdev_op_hold != NULL)
2671 		vd->vdev_ops->vdev_op_hold(vd);
2672 }
2673 
2674 void
2675 vdev_rele(vdev_t *vd)
2676 {
2677 	ASSERT(spa_is_root(vd->vdev_spa));
2678 	for (int c = 0; c < vd->vdev_children; c++)
2679 		vdev_rele(vd->vdev_child[c]);
2680 
2681 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_ops->vdev_op_rele != NULL)
2682 		vd->vdev_ops->vdev_op_rele(vd);
2683 }
2684 
2685 /*
2686  * Reopen all interior vdevs and any unopened leaves.  We don't actually
2687  * reopen leaf vdevs which had previously been opened as they might deadlock
2688  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
2689  * If the leaf has never been opened then open it, as usual.
2690  */
2691 void
2692 vdev_reopen(vdev_t *vd)
2693 {
2694 	spa_t *spa = vd->vdev_spa;
2695 
2696 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2697 
2698 	/* set the reopening flag unless we're taking the vdev offline */
2699 	vd->vdev_reopening = !vd->vdev_offline;
2700 	vdev_close(vd);
2701 	(void) vdev_open(vd);
2702 
2703 	/*
2704 	 * Call vdev_validate() here to make sure we have the same device.
2705 	 * Otherwise, a device with an invalid label could be successfully
2706 	 * opened in response to vdev_reopen().
2707 	 */
2708 	if (vd->vdev_aux) {
2709 		(void) vdev_validate_aux(vd);
2710 		if (vdev_readable(vd) && vdev_writeable(vd) &&
2711 		    vd->vdev_aux == &spa->spa_l2cache) {
2712 			/*
2713 			 * In case the vdev is present we should evict all ARC
2714 			 * buffers and pointers to log blocks and reclaim their
2715 			 * space before restoring its contents to L2ARC.
2716 			 */
2717 			if (l2arc_vdev_present(vd)) {
2718 				l2arc_rebuild_vdev(vd, B_TRUE);
2719 			} else {
2720 				l2arc_add_vdev(spa, vd);
2721 			}
2722 			spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD);
2723 			spa_async_request(spa, SPA_ASYNC_L2CACHE_TRIM);
2724 		}
2725 	} else {
2726 		(void) vdev_validate(vd);
2727 	}
2728 
2729 	/*
2730 	 * Recheck if resilver is still needed and cancel any
2731 	 * scheduled resilver if resilver is unneeded.
2732 	 */
2733 	if (!vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL) &&
2734 	    spa->spa_async_tasks & SPA_ASYNC_RESILVER) {
2735 		mutex_enter(&spa->spa_async_lock);
2736 		spa->spa_async_tasks &= ~SPA_ASYNC_RESILVER;
2737 		mutex_exit(&spa->spa_async_lock);
2738 	}
2739 
2740 	/*
2741 	 * Reassess parent vdev's health.
2742 	 */
2743 	vdev_propagate_state(vd);
2744 }
2745 
2746 int
2747 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
2748 {
2749 	int error;
2750 
2751 	/*
2752 	 * Normally, partial opens (e.g. of a mirror) are allowed.
2753 	 * For a create, however, we want to fail the request if
2754 	 * there are any components we can't open.
2755 	 */
2756 	error = vdev_open(vd);
2757 
2758 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
2759 		vdev_close(vd);
2760 		return (error ? error : SET_ERROR(ENXIO));
2761 	}
2762 
2763 	/*
2764 	 * Recursively load DTLs and initialize all labels.
2765 	 */
2766 	if ((error = vdev_dtl_load(vd)) != 0 ||
2767 	    (error = vdev_label_init(vd, txg, isreplacing ?
2768 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
2769 		vdev_close(vd);
2770 		return (error);
2771 	}
2772 
2773 	return (0);
2774 }
2775 
2776 void
2777 vdev_metaslab_set_size(vdev_t *vd)
2778 {
2779 	uint64_t asize = vd->vdev_asize;
2780 	uint64_t ms_count = asize >> zfs_vdev_default_ms_shift;
2781 	uint64_t ms_shift;
2782 
2783 	/*
2784 	 * There are two dimensions to the metaslab sizing calculation:
2785 	 * the size of the metaslab and the count of metaslabs per vdev.
2786 	 *
2787 	 * The default values used below are a good balance between memory
2788 	 * usage (larger metaslab size means more memory needed for loaded
2789 	 * metaslabs; more metaslabs means more memory needed for the
2790 	 * metaslab_t structs), metaslab load time (larger metaslabs take
2791 	 * longer to load), and metaslab sync time (more metaslabs means
2792 	 * more time spent syncing all of them).
2793 	 *
2794 	 * In general, we aim for zfs_vdev_default_ms_count (200) metaslabs.
2795 	 * The range of the dimensions are as follows:
2796 	 *
2797 	 *	2^29 <= ms_size  <= 2^34
2798 	 *	  16 <= ms_count <= 131,072
2799 	 *
2800 	 * On the lower end of vdev sizes, we aim for metaslabs sizes of
2801 	 * at least 512MB (2^29) to minimize fragmentation effects when
2802 	 * testing with smaller devices.  However, the count constraint
2803 	 * of at least 16 metaslabs will override this minimum size goal.
2804 	 *
2805 	 * On the upper end of vdev sizes, we aim for a maximum metaslab
2806 	 * size of 16GB.  However, we will cap the total count to 2^17
2807 	 * metaslabs to keep our memory footprint in check and let the
2808 	 * metaslab size grow from there if that limit is hit.
2809 	 *
2810 	 * The net effect of applying above constrains is summarized below.
2811 	 *
2812 	 *   vdev size       metaslab count
2813 	 *  --------------|-----------------
2814 	 *      < 8GB        ~16
2815 	 *  8GB   - 100GB   one per 512MB
2816 	 *  100GB - 3TB     ~200
2817 	 *  3TB   - 2PB     one per 16GB
2818 	 *      > 2PB       ~131,072
2819 	 *  --------------------------------
2820 	 *
2821 	 *  Finally, note that all of the above calculate the initial
2822 	 *  number of metaslabs. Expanding a top-level vdev will result
2823 	 *  in additional metaslabs being allocated making it possible
2824 	 *  to exceed the zfs_vdev_ms_count_limit.
2825 	 */
2826 
2827 	if (ms_count < zfs_vdev_min_ms_count)
2828 		ms_shift = highbit64(asize / zfs_vdev_min_ms_count);
2829 	else if (ms_count > zfs_vdev_default_ms_count)
2830 		ms_shift = highbit64(asize / zfs_vdev_default_ms_count);
2831 	else
2832 		ms_shift = zfs_vdev_default_ms_shift;
2833 
2834 	if (ms_shift < SPA_MAXBLOCKSHIFT) {
2835 		ms_shift = SPA_MAXBLOCKSHIFT;
2836 	} else if (ms_shift > zfs_vdev_max_ms_shift) {
2837 		ms_shift = zfs_vdev_max_ms_shift;
2838 		/* cap the total count to constrain memory footprint */
2839 		if ((asize >> ms_shift) > zfs_vdev_ms_count_limit)
2840 			ms_shift = highbit64(asize / zfs_vdev_ms_count_limit);
2841 	}
2842 
2843 	vd->vdev_ms_shift = ms_shift;
2844 	ASSERT3U(vd->vdev_ms_shift, >=, SPA_MAXBLOCKSHIFT);
2845 }
2846 
2847 void
2848 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
2849 {
2850 	ASSERT(vd == vd->vdev_top);
2851 	/* indirect vdevs don't have metaslabs or dtls */
2852 	ASSERT(vdev_is_concrete(vd) || flags == 0);
2853 	ASSERT(ISP2(flags));
2854 	ASSERT(spa_writeable(vd->vdev_spa));
2855 
2856 	if (flags & VDD_METASLAB)
2857 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
2858 
2859 	if (flags & VDD_DTL)
2860 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
2861 
2862 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
2863 }
2864 
2865 void
2866 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
2867 {
2868 	for (int c = 0; c < vd->vdev_children; c++)
2869 		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
2870 
2871 	if (vd->vdev_ops->vdev_op_leaf)
2872 		vdev_dirty(vd->vdev_top, flags, vd, txg);
2873 }
2874 
2875 /*
2876  * DTLs.
2877  *
2878  * A vdev's DTL (dirty time log) is the set of transaction groups for which
2879  * the vdev has less than perfect replication.  There are four kinds of DTL:
2880  *
2881  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
2882  *
2883  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
2884  *
2885  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
2886  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
2887  *	txgs that was scrubbed.
2888  *
2889  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
2890  *	persistent errors or just some device being offline.
2891  *	Unlike the other three, the DTL_OUTAGE map is not generally
2892  *	maintained; it's only computed when needed, typically to
2893  *	determine whether a device can be detached.
2894  *
2895  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
2896  * either has the data or it doesn't.
2897  *
2898  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
2899  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
2900  * if any child is less than fully replicated, then so is its parent.
2901  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
2902  * comprising only those txgs which appear in 'maxfaults' or more children;
2903  * those are the txgs we don't have enough replication to read.  For example,
2904  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
2905  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
2906  * two child DTL_MISSING maps.
2907  *
2908  * It should be clear from the above that to compute the DTLs and outage maps
2909  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
2910  * Therefore, that is all we keep on disk.  When loading the pool, or after
2911  * a configuration change, we generate all other DTLs from first principles.
2912  */
2913 void
2914 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2915 {
2916 	range_tree_t *rt = vd->vdev_dtl[t];
2917 
2918 	ASSERT(t < DTL_TYPES);
2919 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2920 	ASSERT(spa_writeable(vd->vdev_spa));
2921 
2922 	mutex_enter(&vd->vdev_dtl_lock);
2923 	if (!range_tree_contains(rt, txg, size))
2924 		range_tree_add(rt, txg, size);
2925 	mutex_exit(&vd->vdev_dtl_lock);
2926 }
2927 
2928 boolean_t
2929 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2930 {
2931 	range_tree_t *rt = vd->vdev_dtl[t];
2932 	boolean_t dirty = B_FALSE;
2933 
2934 	ASSERT(t < DTL_TYPES);
2935 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2936 
2937 	/*
2938 	 * While we are loading the pool, the DTLs have not been loaded yet.
2939 	 * This isn't a problem but it can result in devices being tried
2940 	 * which are known to not have the data.  In which case, the import
2941 	 * is relying on the checksum to ensure that we get the right data.
2942 	 * Note that while importing we are only reading the MOS, which is
2943 	 * always checksummed.
2944 	 */
2945 	mutex_enter(&vd->vdev_dtl_lock);
2946 	if (!range_tree_is_empty(rt))
2947 		dirty = range_tree_contains(rt, txg, size);
2948 	mutex_exit(&vd->vdev_dtl_lock);
2949 
2950 	return (dirty);
2951 }
2952 
2953 boolean_t
2954 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
2955 {
2956 	range_tree_t *rt = vd->vdev_dtl[t];
2957 	boolean_t empty;
2958 
2959 	mutex_enter(&vd->vdev_dtl_lock);
2960 	empty = range_tree_is_empty(rt);
2961 	mutex_exit(&vd->vdev_dtl_lock);
2962 
2963 	return (empty);
2964 }
2965 
2966 /*
2967  * Check if the txg falls within the range which must be
2968  * resilvered.  DVAs outside this range can always be skipped.
2969  */
2970 boolean_t
2971 vdev_default_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize,
2972     uint64_t phys_birth)
2973 {
2974 	(void) dva, (void) psize;
2975 
2976 	/* Set by sequential resilver. */
2977 	if (phys_birth == TXG_UNKNOWN)
2978 		return (B_TRUE);
2979 
2980 	return (vdev_dtl_contains(vd, DTL_PARTIAL, phys_birth, 1));
2981 }
2982 
2983 /*
2984  * Returns B_TRUE if the vdev determines the DVA needs to be resilvered.
2985  */
2986 boolean_t
2987 vdev_dtl_need_resilver(vdev_t *vd, const dva_t *dva, size_t psize,
2988     uint64_t phys_birth)
2989 {
2990 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2991 
2992 	if (vd->vdev_ops->vdev_op_need_resilver == NULL ||
2993 	    vd->vdev_ops->vdev_op_leaf)
2994 		return (B_TRUE);
2995 
2996 	return (vd->vdev_ops->vdev_op_need_resilver(vd, dva, psize,
2997 	    phys_birth));
2998 }
2999 
3000 /*
3001  * Returns the lowest txg in the DTL range.
3002  */
3003 static uint64_t
3004 vdev_dtl_min(vdev_t *vd)
3005 {
3006 	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
3007 	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
3008 	ASSERT0(vd->vdev_children);
3009 
3010 	return (range_tree_min(vd->vdev_dtl[DTL_MISSING]) - 1);
3011 }
3012 
3013 /*
3014  * Returns the highest txg in the DTL.
3015  */
3016 static uint64_t
3017 vdev_dtl_max(vdev_t *vd)
3018 {
3019 	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
3020 	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
3021 	ASSERT0(vd->vdev_children);
3022 
3023 	return (range_tree_max(vd->vdev_dtl[DTL_MISSING]));
3024 }
3025 
3026 /*
3027  * Determine if a resilvering vdev should remove any DTL entries from
3028  * its range. If the vdev was resilvering for the entire duration of the
3029  * scan then it should excise that range from its DTLs. Otherwise, this
3030  * vdev is considered partially resilvered and should leave its DTL
3031  * entries intact. The comment in vdev_dtl_reassess() describes how we
3032  * excise the DTLs.
3033  */
3034 static boolean_t
3035 vdev_dtl_should_excise(vdev_t *vd, boolean_t rebuild_done)
3036 {
3037 	ASSERT0(vd->vdev_children);
3038 
3039 	if (vd->vdev_state < VDEV_STATE_DEGRADED)
3040 		return (B_FALSE);
3041 
3042 	if (vd->vdev_resilver_deferred)
3043 		return (B_FALSE);
3044 
3045 	if (range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]))
3046 		return (B_TRUE);
3047 
3048 	if (rebuild_done) {
3049 		vdev_rebuild_t *vr = &vd->vdev_top->vdev_rebuild_config;
3050 		vdev_rebuild_phys_t *vrp = &vr->vr_rebuild_phys;
3051 
3052 		/* Rebuild not initiated by attach */
3053 		if (vd->vdev_rebuild_txg == 0)
3054 			return (B_TRUE);
3055 
3056 		/*
3057 		 * When a rebuild completes without error then all missing data
3058 		 * up to the rebuild max txg has been reconstructed and the DTL
3059 		 * is eligible for excision.
3060 		 */
3061 		if (vrp->vrp_rebuild_state == VDEV_REBUILD_COMPLETE &&
3062 		    vdev_dtl_max(vd) <= vrp->vrp_max_txg) {
3063 			ASSERT3U(vrp->vrp_min_txg, <=, vdev_dtl_min(vd));
3064 			ASSERT3U(vrp->vrp_min_txg, <, vd->vdev_rebuild_txg);
3065 			ASSERT3U(vd->vdev_rebuild_txg, <=, vrp->vrp_max_txg);
3066 			return (B_TRUE);
3067 		}
3068 	} else {
3069 		dsl_scan_t *scn = vd->vdev_spa->spa_dsl_pool->dp_scan;
3070 		dsl_scan_phys_t *scnp __maybe_unused = &scn->scn_phys;
3071 
3072 		/* Resilver not initiated by attach */
3073 		if (vd->vdev_resilver_txg == 0)
3074 			return (B_TRUE);
3075 
3076 		/*
3077 		 * When a resilver is initiated the scan will assign the
3078 		 * scn_max_txg value to the highest txg value that exists
3079 		 * in all DTLs. If this device's max DTL is not part of this
3080 		 * scan (i.e. it is not in the range (scn_min_txg, scn_max_txg]
3081 		 * then it is not eligible for excision.
3082 		 */
3083 		if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
3084 			ASSERT3U(scnp->scn_min_txg, <=, vdev_dtl_min(vd));
3085 			ASSERT3U(scnp->scn_min_txg, <, vd->vdev_resilver_txg);
3086 			ASSERT3U(vd->vdev_resilver_txg, <=, scnp->scn_max_txg);
3087 			return (B_TRUE);
3088 		}
3089 	}
3090 
3091 	return (B_FALSE);
3092 }
3093 
3094 /*
3095  * Reassess DTLs after a config change or scrub completion. If txg == 0 no
3096  * write operations will be issued to the pool.
3097  */
3098 void
3099 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg,
3100     boolean_t scrub_done, boolean_t rebuild_done)
3101 {
3102 	spa_t *spa = vd->vdev_spa;
3103 	avl_tree_t reftree;
3104 	int minref;
3105 
3106 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
3107 
3108 	for (int c = 0; c < vd->vdev_children; c++)
3109 		vdev_dtl_reassess(vd->vdev_child[c], txg,
3110 		    scrub_txg, scrub_done, rebuild_done);
3111 
3112 	if (vd == spa->spa_root_vdev || !vdev_is_concrete(vd) || vd->vdev_aux)
3113 		return;
3114 
3115 	if (vd->vdev_ops->vdev_op_leaf) {
3116 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
3117 		vdev_rebuild_t *vr = &vd->vdev_top->vdev_rebuild_config;
3118 		boolean_t check_excise = B_FALSE;
3119 		boolean_t wasempty = B_TRUE;
3120 
3121 		mutex_enter(&vd->vdev_dtl_lock);
3122 
3123 		/*
3124 		 * If requested, pretend the scan or rebuild completed cleanly.
3125 		 */
3126 		if (zfs_scan_ignore_errors) {
3127 			if (scn != NULL)
3128 				scn->scn_phys.scn_errors = 0;
3129 			if (vr != NULL)
3130 				vr->vr_rebuild_phys.vrp_errors = 0;
3131 		}
3132 
3133 		if (scrub_txg != 0 &&
3134 		    !range_tree_is_empty(vd->vdev_dtl[DTL_MISSING])) {
3135 			wasempty = B_FALSE;
3136 			zfs_dbgmsg("guid:%llu txg:%llu scrub:%llu started:%d "
3137 			    "dtl:%llu/%llu errors:%llu",
3138 			    (u_longlong_t)vd->vdev_guid, (u_longlong_t)txg,
3139 			    (u_longlong_t)scrub_txg, spa->spa_scrub_started,
3140 			    (u_longlong_t)vdev_dtl_min(vd),
3141 			    (u_longlong_t)vdev_dtl_max(vd),
3142 			    (u_longlong_t)(scn ? scn->scn_phys.scn_errors : 0));
3143 		}
3144 
3145 		/*
3146 		 * If we've completed a scrub/resilver or a rebuild cleanly
3147 		 * then determine if this vdev should remove any DTLs. We
3148 		 * only want to excise regions on vdevs that were available
3149 		 * during the entire duration of this scan.
3150 		 */
3151 		if (rebuild_done &&
3152 		    vr != NULL && vr->vr_rebuild_phys.vrp_errors == 0) {
3153 			check_excise = B_TRUE;
3154 		} else {
3155 			if (spa->spa_scrub_started ||
3156 			    (scn != NULL && scn->scn_phys.scn_errors == 0)) {
3157 				check_excise = B_TRUE;
3158 			}
3159 		}
3160 
3161 		if (scrub_txg && check_excise &&
3162 		    vdev_dtl_should_excise(vd, rebuild_done)) {
3163 			/*
3164 			 * We completed a scrub, resilver or rebuild up to
3165 			 * scrub_txg.  If we did it without rebooting, then
3166 			 * the scrub dtl will be valid, so excise the old
3167 			 * region and fold in the scrub dtl.  Otherwise,
3168 			 * leave the dtl as-is if there was an error.
3169 			 *
3170 			 * There's little trick here: to excise the beginning
3171 			 * of the DTL_MISSING map, we put it into a reference
3172 			 * tree and then add a segment with refcnt -1 that
3173 			 * covers the range [0, scrub_txg).  This means
3174 			 * that each txg in that range has refcnt -1 or 0.
3175 			 * We then add DTL_SCRUB with a refcnt of 2, so that
3176 			 * entries in the range [0, scrub_txg) will have a
3177 			 * positive refcnt -- either 1 or 2.  We then convert
3178 			 * the reference tree into the new DTL_MISSING map.
3179 			 */
3180 			space_reftree_create(&reftree);
3181 			space_reftree_add_map(&reftree,
3182 			    vd->vdev_dtl[DTL_MISSING], 1);
3183 			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
3184 			space_reftree_add_map(&reftree,
3185 			    vd->vdev_dtl[DTL_SCRUB], 2);
3186 			space_reftree_generate_map(&reftree,
3187 			    vd->vdev_dtl[DTL_MISSING], 1);
3188 			space_reftree_destroy(&reftree);
3189 
3190 			if (!range_tree_is_empty(vd->vdev_dtl[DTL_MISSING])) {
3191 				zfs_dbgmsg("update DTL_MISSING:%llu/%llu",
3192 				    (u_longlong_t)vdev_dtl_min(vd),
3193 				    (u_longlong_t)vdev_dtl_max(vd));
3194 			} else if (!wasempty) {
3195 				zfs_dbgmsg("DTL_MISSING is now empty");
3196 			}
3197 		}
3198 		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
3199 		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
3200 		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
3201 		if (scrub_done)
3202 			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
3203 		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
3204 		if (!vdev_readable(vd))
3205 			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
3206 		else
3207 			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
3208 			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
3209 
3210 		/*
3211 		 * If the vdev was resilvering or rebuilding and no longer
3212 		 * has any DTLs then reset the appropriate flag and dirty
3213 		 * the top level so that we persist the change.
3214 		 */
3215 		if (txg != 0 &&
3216 		    range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) &&
3217 		    range_tree_is_empty(vd->vdev_dtl[DTL_OUTAGE])) {
3218 			if (vd->vdev_rebuild_txg != 0) {
3219 				vd->vdev_rebuild_txg = 0;
3220 				vdev_config_dirty(vd->vdev_top);
3221 			} else if (vd->vdev_resilver_txg != 0) {
3222 				vd->vdev_resilver_txg = 0;
3223 				vdev_config_dirty(vd->vdev_top);
3224 			}
3225 		}
3226 
3227 		mutex_exit(&vd->vdev_dtl_lock);
3228 
3229 		if (txg != 0)
3230 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
3231 		return;
3232 	}
3233 
3234 	mutex_enter(&vd->vdev_dtl_lock);
3235 	for (int t = 0; t < DTL_TYPES; t++) {
3236 		/* account for child's outage in parent's missing map */
3237 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
3238 		if (t == DTL_SCRUB)
3239 			continue;			/* leaf vdevs only */
3240 		if (t == DTL_PARTIAL)
3241 			minref = 1;			/* i.e. non-zero */
3242 		else if (vdev_get_nparity(vd) != 0)
3243 			minref = vdev_get_nparity(vd) + 1; /* RAID-Z, dRAID */
3244 		else
3245 			minref = vd->vdev_children;	/* any kind of mirror */
3246 		space_reftree_create(&reftree);
3247 		for (int c = 0; c < vd->vdev_children; c++) {
3248 			vdev_t *cvd = vd->vdev_child[c];
3249 			mutex_enter(&cvd->vdev_dtl_lock);
3250 			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
3251 			mutex_exit(&cvd->vdev_dtl_lock);
3252 		}
3253 		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
3254 		space_reftree_destroy(&reftree);
3255 	}
3256 	mutex_exit(&vd->vdev_dtl_lock);
3257 }
3258 
3259 /*
3260  * Iterate over all the vdevs except spare, and post kobj events
3261  */
3262 void
3263 vdev_post_kobj_evt(vdev_t *vd)
3264 {
3265 	if (vd->vdev_ops->vdev_op_kobj_evt_post &&
3266 	    vd->vdev_kobj_flag == B_FALSE) {
3267 		vd->vdev_kobj_flag = B_TRUE;
3268 		vd->vdev_ops->vdev_op_kobj_evt_post(vd);
3269 	}
3270 
3271 	for (int c = 0; c < vd->vdev_children; c++)
3272 		vdev_post_kobj_evt(vd->vdev_child[c]);
3273 }
3274 
3275 /*
3276  * Iterate over all the vdevs except spare, and clear kobj events
3277  */
3278 void
3279 vdev_clear_kobj_evt(vdev_t *vd)
3280 {
3281 	vd->vdev_kobj_flag = B_FALSE;
3282 
3283 	for (int c = 0; c < vd->vdev_children; c++)
3284 		vdev_clear_kobj_evt(vd->vdev_child[c]);
3285 }
3286 
3287 int
3288 vdev_dtl_load(vdev_t *vd)
3289 {
3290 	spa_t *spa = vd->vdev_spa;
3291 	objset_t *mos = spa->spa_meta_objset;
3292 	range_tree_t *rt;
3293 	int error = 0;
3294 
3295 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
3296 		ASSERT(vdev_is_concrete(vd));
3297 
3298 		/*
3299 		 * If the dtl cannot be sync'd there is no need to open it.
3300 		 */
3301 		if (spa->spa_mode == SPA_MODE_READ && !spa->spa_read_spacemaps)
3302 			return (0);
3303 
3304 		error = space_map_open(&vd->vdev_dtl_sm, mos,
3305 		    vd->vdev_dtl_object, 0, -1ULL, 0);
3306 		if (error)
3307 			return (error);
3308 		ASSERT(vd->vdev_dtl_sm != NULL);
3309 
3310 		rt = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
3311 		error = space_map_load(vd->vdev_dtl_sm, rt, SM_ALLOC);
3312 		if (error == 0) {
3313 			mutex_enter(&vd->vdev_dtl_lock);
3314 			range_tree_walk(rt, range_tree_add,
3315 			    vd->vdev_dtl[DTL_MISSING]);
3316 			mutex_exit(&vd->vdev_dtl_lock);
3317 		}
3318 
3319 		range_tree_vacate(rt, NULL, NULL);
3320 		range_tree_destroy(rt);
3321 
3322 		return (error);
3323 	}
3324 
3325 	for (int c = 0; c < vd->vdev_children; c++) {
3326 		error = vdev_dtl_load(vd->vdev_child[c]);
3327 		if (error != 0)
3328 			break;
3329 	}
3330 
3331 	return (error);
3332 }
3333 
3334 static void
3335 vdev_zap_allocation_data(vdev_t *vd, dmu_tx_t *tx)
3336 {
3337 	spa_t *spa = vd->vdev_spa;
3338 	objset_t *mos = spa->spa_meta_objset;
3339 	vdev_alloc_bias_t alloc_bias = vd->vdev_alloc_bias;
3340 	const char *string;
3341 
3342 	ASSERT(alloc_bias != VDEV_BIAS_NONE);
3343 
3344 	string =
3345 	    (alloc_bias == VDEV_BIAS_LOG) ? VDEV_ALLOC_BIAS_LOG :
3346 	    (alloc_bias == VDEV_BIAS_SPECIAL) ? VDEV_ALLOC_BIAS_SPECIAL :
3347 	    (alloc_bias == VDEV_BIAS_DEDUP) ? VDEV_ALLOC_BIAS_DEDUP : NULL;
3348 
3349 	ASSERT(string != NULL);
3350 	VERIFY0(zap_add(mos, vd->vdev_top_zap, VDEV_TOP_ZAP_ALLOCATION_BIAS,
3351 	    1, strlen(string) + 1, string, tx));
3352 
3353 	if (alloc_bias == VDEV_BIAS_SPECIAL || alloc_bias == VDEV_BIAS_DEDUP) {
3354 		spa_activate_allocation_classes(spa, tx);
3355 	}
3356 }
3357 
3358 void
3359 vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
3360 {
3361 	spa_t *spa = vd->vdev_spa;
3362 
3363 	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
3364 	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
3365 	    zapobj, tx));
3366 }
3367 
3368 uint64_t
3369 vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
3370 {
3371 	spa_t *spa = vd->vdev_spa;
3372 	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
3373 	    DMU_OT_NONE, 0, tx);
3374 
3375 	ASSERT(zap != 0);
3376 	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
3377 	    zap, tx));
3378 
3379 	return (zap);
3380 }
3381 
3382 void
3383 vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
3384 {
3385 	if (vd->vdev_ops != &vdev_hole_ops &&
3386 	    vd->vdev_ops != &vdev_missing_ops &&
3387 	    vd->vdev_ops != &vdev_root_ops &&
3388 	    !vd->vdev_top->vdev_removing) {
3389 		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
3390 			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
3391 		}
3392 		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
3393 			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
3394 			if (vd->vdev_alloc_bias != VDEV_BIAS_NONE)
3395 				vdev_zap_allocation_data(vd, tx);
3396 		}
3397 	}
3398 	if (vd->vdev_ops == &vdev_root_ops && vd->vdev_root_zap == 0 &&
3399 	    spa_feature_is_enabled(vd->vdev_spa, SPA_FEATURE_AVZ_V2)) {
3400 		if (!spa_feature_is_active(vd->vdev_spa, SPA_FEATURE_AVZ_V2))
3401 			spa_feature_incr(vd->vdev_spa, SPA_FEATURE_AVZ_V2, tx);
3402 		vd->vdev_root_zap = vdev_create_link_zap(vd, tx);
3403 	}
3404 
3405 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
3406 		vdev_construct_zaps(vd->vdev_child[i], tx);
3407 	}
3408 }
3409 
3410 static void
3411 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
3412 {
3413 	spa_t *spa = vd->vdev_spa;
3414 	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
3415 	objset_t *mos = spa->spa_meta_objset;
3416 	range_tree_t *rtsync;
3417 	dmu_tx_t *tx;
3418 	uint64_t object = space_map_object(vd->vdev_dtl_sm);
3419 
3420 	ASSERT(vdev_is_concrete(vd));
3421 	ASSERT(vd->vdev_ops->vdev_op_leaf);
3422 
3423 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
3424 
3425 	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
3426 		mutex_enter(&vd->vdev_dtl_lock);
3427 		space_map_free(vd->vdev_dtl_sm, tx);
3428 		space_map_close(vd->vdev_dtl_sm);
3429 		vd->vdev_dtl_sm = NULL;
3430 		mutex_exit(&vd->vdev_dtl_lock);
3431 
3432 		/*
3433 		 * We only destroy the leaf ZAP for detached leaves or for
3434 		 * removed log devices. Removed data devices handle leaf ZAP
3435 		 * cleanup later, once cancellation is no longer possible.
3436 		 */
3437 		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
3438 		    vd->vdev_top->vdev_islog)) {
3439 			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
3440 			vd->vdev_leaf_zap = 0;
3441 		}
3442 
3443 		dmu_tx_commit(tx);
3444 		return;
3445 	}
3446 
3447 	if (vd->vdev_dtl_sm == NULL) {
3448 		uint64_t new_object;
3449 
3450 		new_object = space_map_alloc(mos, zfs_vdev_dtl_sm_blksz, tx);
3451 		VERIFY3U(new_object, !=, 0);
3452 
3453 		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
3454 		    0, -1ULL, 0));
3455 		ASSERT(vd->vdev_dtl_sm != NULL);
3456 	}
3457 
3458 	rtsync = range_tree_create(NULL, RANGE_SEG64, NULL, 0, 0);
3459 
3460 	mutex_enter(&vd->vdev_dtl_lock);
3461 	range_tree_walk(rt, range_tree_add, rtsync);
3462 	mutex_exit(&vd->vdev_dtl_lock);
3463 
3464 	space_map_truncate(vd->vdev_dtl_sm, zfs_vdev_dtl_sm_blksz, tx);
3465 	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, SM_NO_VDEVID, tx);
3466 	range_tree_vacate(rtsync, NULL, NULL);
3467 
3468 	range_tree_destroy(rtsync);
3469 
3470 	/*
3471 	 * If the object for the space map has changed then dirty
3472 	 * the top level so that we update the config.
3473 	 */
3474 	if (object != space_map_object(vd->vdev_dtl_sm)) {
3475 		vdev_dbgmsg(vd, "txg %llu, spa %s, DTL old object %llu, "
3476 		    "new object %llu", (u_longlong_t)txg, spa_name(spa),
3477 		    (u_longlong_t)object,
3478 		    (u_longlong_t)space_map_object(vd->vdev_dtl_sm));
3479 		vdev_config_dirty(vd->vdev_top);
3480 	}
3481 
3482 	dmu_tx_commit(tx);
3483 }
3484 
3485 /*
3486  * Determine whether the specified vdev can be offlined/detached/removed
3487  * without losing data.
3488  */
3489 boolean_t
3490 vdev_dtl_required(vdev_t *vd)
3491 {
3492 	spa_t *spa = vd->vdev_spa;
3493 	vdev_t *tvd = vd->vdev_top;
3494 	uint8_t cant_read = vd->vdev_cant_read;
3495 	boolean_t required;
3496 
3497 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3498 
3499 	if (vd == spa->spa_root_vdev || vd == tvd)
3500 		return (B_TRUE);
3501 
3502 	/*
3503 	 * Temporarily mark the device as unreadable, and then determine
3504 	 * whether this results in any DTL outages in the top-level vdev.
3505 	 * If not, we can safely offline/detach/remove the device.
3506 	 */
3507 	vd->vdev_cant_read = B_TRUE;
3508 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE, B_FALSE);
3509 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
3510 	vd->vdev_cant_read = cant_read;
3511 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE, B_FALSE);
3512 
3513 	if (!required && zio_injection_enabled) {
3514 		required = !!zio_handle_device_injection(vd, NULL,
3515 		    SET_ERROR(ECHILD));
3516 	}
3517 
3518 	return (required);
3519 }
3520 
3521 /*
3522  * Determine if resilver is needed, and if so the txg range.
3523  */
3524 boolean_t
3525 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
3526 {
3527 	boolean_t needed = B_FALSE;
3528 	uint64_t thismin = UINT64_MAX;
3529 	uint64_t thismax = 0;
3530 
3531 	if (vd->vdev_children == 0) {
3532 		mutex_enter(&vd->vdev_dtl_lock);
3533 		if (!range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) &&
3534 		    vdev_writeable(vd)) {
3535 
3536 			thismin = vdev_dtl_min(vd);
3537 			thismax = vdev_dtl_max(vd);
3538 			needed = B_TRUE;
3539 		}
3540 		mutex_exit(&vd->vdev_dtl_lock);
3541 	} else {
3542 		for (int c = 0; c < vd->vdev_children; c++) {
3543 			vdev_t *cvd = vd->vdev_child[c];
3544 			uint64_t cmin, cmax;
3545 
3546 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
3547 				thismin = MIN(thismin, cmin);
3548 				thismax = MAX(thismax, cmax);
3549 				needed = B_TRUE;
3550 			}
3551 		}
3552 	}
3553 
3554 	if (needed && minp) {
3555 		*minp = thismin;
3556 		*maxp = thismax;
3557 	}
3558 	return (needed);
3559 }
3560 
3561 /*
3562  * Gets the checkpoint space map object from the vdev's ZAP.  On success sm_obj
3563  * will contain either the checkpoint spacemap object or zero if none exists.
3564  * All other errors are returned to the caller.
3565  */
3566 int
3567 vdev_checkpoint_sm_object(vdev_t *vd, uint64_t *sm_obj)
3568 {
3569 	ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
3570 
3571 	if (vd->vdev_top_zap == 0) {
3572 		*sm_obj = 0;
3573 		return (0);
3574 	}
3575 
3576 	int error = zap_lookup(spa_meta_objset(vd->vdev_spa), vd->vdev_top_zap,
3577 	    VDEV_TOP_ZAP_POOL_CHECKPOINT_SM, sizeof (uint64_t), 1, sm_obj);
3578 	if (error == ENOENT) {
3579 		*sm_obj = 0;
3580 		error = 0;
3581 	}
3582 
3583 	return (error);
3584 }
3585 
3586 int
3587 vdev_load(vdev_t *vd)
3588 {
3589 	int children = vd->vdev_children;
3590 	int error = 0;
3591 	taskq_t *tq = NULL;
3592 
3593 	/*
3594 	 * It's only worthwhile to use the taskq for the root vdev, because the
3595 	 * slow part is metaslab_init, and that only happens for top-level
3596 	 * vdevs.
3597 	 */
3598 	if (vd->vdev_ops == &vdev_root_ops && vd->vdev_children > 0) {
3599 		tq = taskq_create("vdev_load", children, minclsyspri,
3600 		    children, children, TASKQ_PREPOPULATE);
3601 	}
3602 
3603 	/*
3604 	 * Recursively load all children.
3605 	 */
3606 	for (int c = 0; c < vd->vdev_children; c++) {
3607 		vdev_t *cvd = vd->vdev_child[c];
3608 
3609 		if (tq == NULL || vdev_uses_zvols(cvd)) {
3610 			cvd->vdev_load_error = vdev_load(cvd);
3611 		} else {
3612 			VERIFY(taskq_dispatch(tq, vdev_load_child,
3613 			    cvd, TQ_SLEEP) != TASKQID_INVALID);
3614 		}
3615 	}
3616 
3617 	if (tq != NULL) {
3618 		taskq_wait(tq);
3619 		taskq_destroy(tq);
3620 	}
3621 
3622 	for (int c = 0; c < vd->vdev_children; c++) {
3623 		int error = vd->vdev_child[c]->vdev_load_error;
3624 
3625 		if (error != 0)
3626 			return (error);
3627 	}
3628 
3629 	vdev_set_deflate_ratio(vd);
3630 
3631 	/*
3632 	 * On spa_load path, grab the allocation bias from our zap
3633 	 */
3634 	if (vd == vd->vdev_top && vd->vdev_top_zap != 0) {
3635 		spa_t *spa = vd->vdev_spa;
3636 		char bias_str[64];
3637 
3638 		error = zap_lookup(spa->spa_meta_objset, vd->vdev_top_zap,
3639 		    VDEV_TOP_ZAP_ALLOCATION_BIAS, 1, sizeof (bias_str),
3640 		    bias_str);
3641 		if (error == 0) {
3642 			ASSERT(vd->vdev_alloc_bias == VDEV_BIAS_NONE);
3643 			vd->vdev_alloc_bias = vdev_derive_alloc_bias(bias_str);
3644 		} else if (error != ENOENT) {
3645 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3646 			    VDEV_AUX_CORRUPT_DATA);
3647 			vdev_dbgmsg(vd, "vdev_load: zap_lookup(top_zap=%llu) "
3648 			    "failed [error=%d]",
3649 			    (u_longlong_t)vd->vdev_top_zap, error);
3650 			return (error);
3651 		}
3652 	}
3653 
3654 	if (vd == vd->vdev_top && vd->vdev_top_zap != 0) {
3655 		spa_t *spa = vd->vdev_spa;
3656 		uint64_t failfast;
3657 
3658 		error = zap_lookup(spa->spa_meta_objset, vd->vdev_top_zap,
3659 		    vdev_prop_to_name(VDEV_PROP_FAILFAST), sizeof (failfast),
3660 		    1, &failfast);
3661 		if (error == 0) {
3662 			vd->vdev_failfast = failfast & 1;
3663 		} else if (error == ENOENT) {
3664 			vd->vdev_failfast = vdev_prop_default_numeric(
3665 			    VDEV_PROP_FAILFAST);
3666 		} else {
3667 			vdev_dbgmsg(vd,
3668 			    "vdev_load: zap_lookup(top_zap=%llu) "
3669 			    "failed [error=%d]",
3670 			    (u_longlong_t)vd->vdev_top_zap, error);
3671 		}
3672 	}
3673 
3674 	/*
3675 	 * Load any rebuild state from the top-level vdev zap.
3676 	 */
3677 	if (vd == vd->vdev_top && vd->vdev_top_zap != 0) {
3678 		error = vdev_rebuild_load(vd);
3679 		if (error && error != ENOTSUP) {
3680 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3681 			    VDEV_AUX_CORRUPT_DATA);
3682 			vdev_dbgmsg(vd, "vdev_load: vdev_rebuild_load "
3683 			    "failed [error=%d]", error);
3684 			return (error);
3685 		}
3686 	}
3687 
3688 	if (vd->vdev_top_zap != 0 || vd->vdev_leaf_zap != 0) {
3689 		uint64_t zapobj;
3690 
3691 		if (vd->vdev_top_zap != 0)
3692 			zapobj = vd->vdev_top_zap;
3693 		else
3694 			zapobj = vd->vdev_leaf_zap;
3695 
3696 		error = vdev_prop_get_int(vd, VDEV_PROP_CHECKSUM_N,
3697 		    &vd->vdev_checksum_n);
3698 		if (error && error != ENOENT)
3699 			vdev_dbgmsg(vd, "vdev_load: zap_lookup(zap=%llu) "
3700 			    "failed [error=%d]", (u_longlong_t)zapobj, error);
3701 
3702 		error = vdev_prop_get_int(vd, VDEV_PROP_CHECKSUM_T,
3703 		    &vd->vdev_checksum_t);
3704 		if (error && error != ENOENT)
3705 			vdev_dbgmsg(vd, "vdev_load: zap_lookup(zap=%llu) "
3706 			    "failed [error=%d]", (u_longlong_t)zapobj, error);
3707 
3708 		error = vdev_prop_get_int(vd, VDEV_PROP_IO_N,
3709 		    &vd->vdev_io_n);
3710 		if (error && error != ENOENT)
3711 			vdev_dbgmsg(vd, "vdev_load: zap_lookup(zap=%llu) "
3712 			    "failed [error=%d]", (u_longlong_t)zapobj, error);
3713 
3714 		error = vdev_prop_get_int(vd, VDEV_PROP_IO_T,
3715 		    &vd->vdev_io_t);
3716 		if (error && error != ENOENT)
3717 			vdev_dbgmsg(vd, "vdev_load: zap_lookup(zap=%llu) "
3718 			    "failed [error=%d]", (u_longlong_t)zapobj, error);
3719 	}
3720 
3721 	/*
3722 	 * If this is a top-level vdev, initialize its metaslabs.
3723 	 */
3724 	if (vd == vd->vdev_top && vdev_is_concrete(vd)) {
3725 		vdev_metaslab_group_create(vd);
3726 
3727 		if (vd->vdev_ashift == 0 || vd->vdev_asize == 0) {
3728 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3729 			    VDEV_AUX_CORRUPT_DATA);
3730 			vdev_dbgmsg(vd, "vdev_load: invalid size. ashift=%llu, "
3731 			    "asize=%llu", (u_longlong_t)vd->vdev_ashift,
3732 			    (u_longlong_t)vd->vdev_asize);
3733 			return (SET_ERROR(ENXIO));
3734 		}
3735 
3736 		error = vdev_metaslab_init(vd, 0);
3737 		if (error != 0) {
3738 			vdev_dbgmsg(vd, "vdev_load: metaslab_init failed "
3739 			    "[error=%d]", error);
3740 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3741 			    VDEV_AUX_CORRUPT_DATA);
3742 			return (error);
3743 		}
3744 
3745 		uint64_t checkpoint_sm_obj;
3746 		error = vdev_checkpoint_sm_object(vd, &checkpoint_sm_obj);
3747 		if (error == 0 && checkpoint_sm_obj != 0) {
3748 			objset_t *mos = spa_meta_objset(vd->vdev_spa);
3749 			ASSERT(vd->vdev_asize != 0);
3750 			ASSERT3P(vd->vdev_checkpoint_sm, ==, NULL);
3751 
3752 			error = space_map_open(&vd->vdev_checkpoint_sm,
3753 			    mos, checkpoint_sm_obj, 0, vd->vdev_asize,
3754 			    vd->vdev_ashift);
3755 			if (error != 0) {
3756 				vdev_dbgmsg(vd, "vdev_load: space_map_open "
3757 				    "failed for checkpoint spacemap (obj %llu) "
3758 				    "[error=%d]",
3759 				    (u_longlong_t)checkpoint_sm_obj, error);
3760 				return (error);
3761 			}
3762 			ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL);
3763 
3764 			/*
3765 			 * Since the checkpoint_sm contains free entries
3766 			 * exclusively we can use space_map_allocated() to
3767 			 * indicate the cumulative checkpointed space that
3768 			 * has been freed.
3769 			 */
3770 			vd->vdev_stat.vs_checkpoint_space =
3771 			    -space_map_allocated(vd->vdev_checkpoint_sm);
3772 			vd->vdev_spa->spa_checkpoint_info.sci_dspace +=
3773 			    vd->vdev_stat.vs_checkpoint_space;
3774 		} else if (error != 0) {
3775 			vdev_dbgmsg(vd, "vdev_load: failed to retrieve "
3776 			    "checkpoint space map object from vdev ZAP "
3777 			    "[error=%d]", error);
3778 			return (error);
3779 		}
3780 	}
3781 
3782 	/*
3783 	 * If this is a leaf vdev, load its DTL.
3784 	 */
3785 	if (vd->vdev_ops->vdev_op_leaf && (error = vdev_dtl_load(vd)) != 0) {
3786 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3787 		    VDEV_AUX_CORRUPT_DATA);
3788 		vdev_dbgmsg(vd, "vdev_load: vdev_dtl_load failed "
3789 		    "[error=%d]", error);
3790 		return (error);
3791 	}
3792 
3793 	uint64_t obsolete_sm_object;
3794 	error = vdev_obsolete_sm_object(vd, &obsolete_sm_object);
3795 	if (error == 0 && obsolete_sm_object != 0) {
3796 		objset_t *mos = vd->vdev_spa->spa_meta_objset;
3797 		ASSERT(vd->vdev_asize != 0);
3798 		ASSERT3P(vd->vdev_obsolete_sm, ==, NULL);
3799 
3800 		if ((error = space_map_open(&vd->vdev_obsolete_sm, mos,
3801 		    obsolete_sm_object, 0, vd->vdev_asize, 0))) {
3802 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
3803 			    VDEV_AUX_CORRUPT_DATA);
3804 			vdev_dbgmsg(vd, "vdev_load: space_map_open failed for "
3805 			    "obsolete spacemap (obj %llu) [error=%d]",
3806 			    (u_longlong_t)obsolete_sm_object, error);
3807 			return (error);
3808 		}
3809 	} else if (error != 0) {
3810 		vdev_dbgmsg(vd, "vdev_load: failed to retrieve obsolete "
3811 		    "space map object from vdev ZAP [error=%d]", error);
3812 		return (error);
3813 	}
3814 
3815 	return (0);
3816 }
3817 
3818 /*
3819  * The special vdev case is used for hot spares and l2cache devices.  Its
3820  * sole purpose it to set the vdev state for the associated vdev.  To do this,
3821  * we make sure that we can open the underlying device, then try to read the
3822  * label, and make sure that the label is sane and that it hasn't been
3823  * repurposed to another pool.
3824  */
3825 int
3826 vdev_validate_aux(vdev_t *vd)
3827 {
3828 	nvlist_t *label;
3829 	uint64_t guid, version;
3830 	uint64_t state;
3831 
3832 	if (!vdev_readable(vd))
3833 		return (0);
3834 
3835 	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
3836 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
3837 		    VDEV_AUX_CORRUPT_DATA);
3838 		return (-1);
3839 	}
3840 
3841 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
3842 	    !SPA_VERSION_IS_SUPPORTED(version) ||
3843 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
3844 	    guid != vd->vdev_guid ||
3845 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
3846 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
3847 		    VDEV_AUX_CORRUPT_DATA);
3848 		nvlist_free(label);
3849 		return (-1);
3850 	}
3851 
3852 	/*
3853 	 * We don't actually check the pool state here.  If it's in fact in
3854 	 * use by another pool, we update this fact on the fly when requested.
3855 	 */
3856 	nvlist_free(label);
3857 	return (0);
3858 }
3859 
3860 static void
3861 vdev_destroy_ms_flush_data(vdev_t *vd, dmu_tx_t *tx)
3862 {
3863 	objset_t *mos = spa_meta_objset(vd->vdev_spa);
3864 
3865 	if (vd->vdev_top_zap == 0)
3866 		return;
3867 
3868 	uint64_t object = 0;
3869 	int err = zap_lookup(mos, vd->vdev_top_zap,
3870 	    VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, sizeof (uint64_t), 1, &object);
3871 	if (err == ENOENT)
3872 		return;
3873 	VERIFY0(err);
3874 
3875 	VERIFY0(dmu_object_free(mos, object, tx));
3876 	VERIFY0(zap_remove(mos, vd->vdev_top_zap,
3877 	    VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS, tx));
3878 }
3879 
3880 /*
3881  * Free the objects used to store this vdev's spacemaps, and the array
3882  * that points to them.
3883  */
3884 void
3885 vdev_destroy_spacemaps(vdev_t *vd, dmu_tx_t *tx)
3886 {
3887 	if (vd->vdev_ms_array == 0)
3888 		return;
3889 
3890 	objset_t *mos = vd->vdev_spa->spa_meta_objset;
3891 	uint64_t array_count = vd->vdev_asize >> vd->vdev_ms_shift;
3892 	size_t array_bytes = array_count * sizeof (uint64_t);
3893 	uint64_t *smobj_array = kmem_alloc(array_bytes, KM_SLEEP);
3894 	VERIFY0(dmu_read(mos, vd->vdev_ms_array, 0,
3895 	    array_bytes, smobj_array, 0));
3896 
3897 	for (uint64_t i = 0; i < array_count; i++) {
3898 		uint64_t smobj = smobj_array[i];
3899 		if (smobj == 0)
3900 			continue;
3901 
3902 		space_map_free_obj(mos, smobj, tx);
3903 	}
3904 
3905 	kmem_free(smobj_array, array_bytes);
3906 	VERIFY0(dmu_object_free(mos, vd->vdev_ms_array, tx));
3907 	vdev_destroy_ms_flush_data(vd, tx);
3908 	vd->vdev_ms_array = 0;
3909 }
3910 
3911 static void
3912 vdev_remove_empty_log(vdev_t *vd, uint64_t txg)
3913 {
3914 	spa_t *spa = vd->vdev_spa;
3915 
3916 	ASSERT(vd->vdev_islog);
3917 	ASSERT(vd == vd->vdev_top);
3918 	ASSERT3U(txg, ==, spa_syncing_txg(spa));
3919 
3920 	dmu_tx_t *tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
3921 
3922 	vdev_destroy_spacemaps(vd, tx);
3923 	if (vd->vdev_top_zap != 0) {
3924 		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
3925 		vd->vdev_top_zap = 0;
3926 	}
3927 
3928 	dmu_tx_commit(tx);
3929 }
3930 
3931 void
3932 vdev_sync_done(vdev_t *vd, uint64_t txg)
3933 {
3934 	metaslab_t *msp;
3935 	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
3936 
3937 	ASSERT(vdev_is_concrete(vd));
3938 
3939 	while ((msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
3940 	    != NULL)
3941 		metaslab_sync_done(msp, txg);
3942 
3943 	if (reassess) {
3944 		metaslab_sync_reassess(vd->vdev_mg);
3945 		if (vd->vdev_log_mg != NULL)
3946 			metaslab_sync_reassess(vd->vdev_log_mg);
3947 	}
3948 }
3949 
3950 void
3951 vdev_sync(vdev_t *vd, uint64_t txg)
3952 {
3953 	spa_t *spa = vd->vdev_spa;
3954 	vdev_t *lvd;
3955 	metaslab_t *msp;
3956 
3957 	ASSERT3U(txg, ==, spa->spa_syncing_txg);
3958 	dmu_tx_t *tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
3959 	if (range_tree_space(vd->vdev_obsolete_segments) > 0) {
3960 		ASSERT(vd->vdev_removing ||
3961 		    vd->vdev_ops == &vdev_indirect_ops);
3962 
3963 		vdev_indirect_sync_obsolete(vd, tx);
3964 
3965 		/*
3966 		 * If the vdev is indirect, it can't have dirty
3967 		 * metaslabs or DTLs.
3968 		 */
3969 		if (vd->vdev_ops == &vdev_indirect_ops) {
3970 			ASSERT(txg_list_empty(&vd->vdev_ms_list, txg));
3971 			ASSERT(txg_list_empty(&vd->vdev_dtl_list, txg));
3972 			dmu_tx_commit(tx);
3973 			return;
3974 		}
3975 	}
3976 
3977 	ASSERT(vdev_is_concrete(vd));
3978 
3979 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0 &&
3980 	    !vd->vdev_removing) {
3981 		ASSERT(vd == vd->vdev_top);
3982 		ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
3983 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
3984 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
3985 		ASSERT(vd->vdev_ms_array != 0);
3986 		vdev_config_dirty(vd);
3987 	}
3988 
3989 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
3990 		metaslab_sync(msp, txg);
3991 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
3992 	}
3993 
3994 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
3995 		vdev_dtl_sync(lvd, txg);
3996 
3997 	/*
3998 	 * If this is an empty log device being removed, destroy the
3999 	 * metadata associated with it.
4000 	 */
4001 	if (vd->vdev_islog && vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
4002 		vdev_remove_empty_log(vd, txg);
4003 
4004 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
4005 	dmu_tx_commit(tx);
4006 }
4007 
4008 uint64_t
4009 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
4010 {
4011 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
4012 }
4013 
4014 /*
4015  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
4016  * not be opened, and no I/O is attempted.
4017  */
4018 int
4019 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
4020 {
4021 	vdev_t *vd, *tvd;
4022 
4023 	spa_vdev_state_enter(spa, SCL_NONE);
4024 
4025 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
4026 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
4027 
4028 	if (!vd->vdev_ops->vdev_op_leaf)
4029 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENOTSUP)));
4030 
4031 	tvd = vd->vdev_top;
4032 
4033 	/*
4034 	 * If user did a 'zpool offline -f' then make the fault persist across
4035 	 * reboots.
4036 	 */
4037 	if (aux == VDEV_AUX_EXTERNAL_PERSIST) {
4038 		/*
4039 		 * There are two kinds of forced faults: temporary and
4040 		 * persistent.  Temporary faults go away at pool import, while
4041 		 * persistent faults stay set.  Both types of faults can be
4042 		 * cleared with a zpool clear.
4043 		 *
4044 		 * We tell if a vdev is persistently faulted by looking at the
4045 		 * ZPOOL_CONFIG_AUX_STATE nvpair.  If it's set to "external" at
4046 		 * import then it's a persistent fault.  Otherwise, it's
4047 		 * temporary.  We get ZPOOL_CONFIG_AUX_STATE set to "external"
4048 		 * by setting vd.vdev_stat.vs_aux to VDEV_AUX_EXTERNAL.  This
4049 		 * tells vdev_config_generate() (which gets run later) to set
4050 		 * ZPOOL_CONFIG_AUX_STATE to "external" in the nvlist.
4051 		 */
4052 		vd->vdev_stat.vs_aux = VDEV_AUX_EXTERNAL;
4053 		vd->vdev_tmpoffline = B_FALSE;
4054 		aux = VDEV_AUX_EXTERNAL;
4055 	} else {
4056 		vd->vdev_tmpoffline = B_TRUE;
4057 	}
4058 
4059 	/*
4060 	 * We don't directly use the aux state here, but if we do a
4061 	 * vdev_reopen(), we need this value to be present to remember why we
4062 	 * were faulted.
4063 	 */
4064 	vd->vdev_label_aux = aux;
4065 
4066 	/*
4067 	 * Faulted state takes precedence over degraded.
4068 	 */
4069 	vd->vdev_delayed_close = B_FALSE;
4070 	vd->vdev_faulted = 1ULL;
4071 	vd->vdev_degraded = 0ULL;
4072 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
4073 
4074 	/*
4075 	 * If this device has the only valid copy of the data, then
4076 	 * back off and simply mark the vdev as degraded instead.
4077 	 */
4078 	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
4079 		vd->vdev_degraded = 1ULL;
4080 		vd->vdev_faulted = 0ULL;
4081 
4082 		/*
4083 		 * If we reopen the device and it's not dead, only then do we
4084 		 * mark it degraded.
4085 		 */
4086 		vdev_reopen(tvd);
4087 
4088 		if (vdev_readable(vd))
4089 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
4090 	}
4091 
4092 	return (spa_vdev_state_exit(spa, vd, 0));
4093 }
4094 
4095 /*
4096  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
4097  * user that something is wrong.  The vdev continues to operate as normal as far
4098  * as I/O is concerned.
4099  */
4100 int
4101 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
4102 {
4103 	vdev_t *vd;
4104 
4105 	spa_vdev_state_enter(spa, SCL_NONE);
4106 
4107 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
4108 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
4109 
4110 	if (!vd->vdev_ops->vdev_op_leaf)
4111 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENOTSUP)));
4112 
4113 	/*
4114 	 * If the vdev is already faulted, then don't do anything.
4115 	 */
4116 	if (vd->vdev_faulted || vd->vdev_degraded)
4117 		return (spa_vdev_state_exit(spa, NULL, 0));
4118 
4119 	vd->vdev_degraded = 1ULL;
4120 	if (!vdev_is_dead(vd))
4121 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
4122 		    aux);
4123 
4124 	return (spa_vdev_state_exit(spa, vd, 0));
4125 }
4126 
4127 int
4128 vdev_remove_wanted(spa_t *spa, uint64_t guid)
4129 {
4130 	vdev_t *vd;
4131 
4132 	spa_vdev_state_enter(spa, SCL_NONE);
4133 
4134 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
4135 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
4136 
4137 	/*
4138 	 * If the vdev is already removed, or expanding which can trigger
4139 	 * repartition add/remove events, then don't do anything.
4140 	 */
4141 	if (vd->vdev_removed || vd->vdev_expanding)
4142 		return (spa_vdev_state_exit(spa, NULL, 0));
4143 
4144 	/*
4145 	 * Confirm the vdev has been removed, otherwise don't do anything.
4146 	 */
4147 	if (vd->vdev_ops->vdev_op_leaf && !zio_wait(vdev_probe(vd, NULL)))
4148 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(EEXIST)));
4149 
4150 	vd->vdev_remove_wanted = B_TRUE;
4151 	spa_async_request(spa, SPA_ASYNC_REMOVE);
4152 
4153 	return (spa_vdev_state_exit(spa, vd, 0));
4154 }
4155 
4156 
4157 /*
4158  * Online the given vdev.
4159  *
4160  * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
4161  * spare device should be detached when the device finishes resilvering.
4162  * Second, the online should be treated like a 'test' online case, so no FMA
4163  * events are generated if the device fails to open.
4164  */
4165 int
4166 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
4167 {
4168 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
4169 	boolean_t wasoffline;
4170 	vdev_state_t oldstate;
4171 
4172 	spa_vdev_state_enter(spa, SCL_NONE);
4173 
4174 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
4175 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
4176 
4177 	if (!vd->vdev_ops->vdev_op_leaf)
4178 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENOTSUP)));
4179 
4180 	wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
4181 	oldstate = vd->vdev_state;
4182 
4183 	tvd = vd->vdev_top;
4184 	vd->vdev_offline = B_FALSE;
4185 	vd->vdev_tmpoffline = B_FALSE;
4186 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
4187 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
4188 
4189 	/* XXX - L2ARC 1.0 does not support expansion */
4190 	if (!vd->vdev_aux) {
4191 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
4192 			pvd->vdev_expanding = !!((flags & ZFS_ONLINE_EXPAND) ||
4193 			    spa->spa_autoexpand);
4194 		vd->vdev_expansion_time = gethrestime_sec();
4195 	}
4196 
4197 	vdev_reopen(tvd);
4198 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
4199 
4200 	if (!vd->vdev_aux) {
4201 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
4202 			pvd->vdev_expanding = B_FALSE;
4203 	}
4204 
4205 	if (newstate)
4206 		*newstate = vd->vdev_state;
4207 	if ((flags & ZFS_ONLINE_UNSPARE) &&
4208 	    !vdev_is_dead(vd) && vd->vdev_parent &&
4209 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
4210 	    vd->vdev_parent->vdev_child[0] == vd)
4211 		vd->vdev_unspare = B_TRUE;
4212 
4213 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
4214 
4215 		/* XXX - L2ARC 1.0 does not support expansion */
4216 		if (vd->vdev_aux)
4217 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
4218 		spa->spa_ccw_fail_time = 0;
4219 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
4220 	}
4221 
4222 	/* Restart initializing if necessary */
4223 	mutex_enter(&vd->vdev_initialize_lock);
4224 	if (vdev_writeable(vd) &&
4225 	    vd->vdev_initialize_thread == NULL &&
4226 	    vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) {
4227 		(void) vdev_initialize(vd);
4228 	}
4229 	mutex_exit(&vd->vdev_initialize_lock);
4230 
4231 	/*
4232 	 * Restart trimming if necessary. We do not restart trimming for cache
4233 	 * devices here. This is triggered by l2arc_rebuild_vdev()
4234 	 * asynchronously for the whole device or in l2arc_evict() as it evicts
4235 	 * space for upcoming writes.
4236 	 */
4237 	mutex_enter(&vd->vdev_trim_lock);
4238 	if (vdev_writeable(vd) && !vd->vdev_isl2cache &&
4239 	    vd->vdev_trim_thread == NULL &&
4240 	    vd->vdev_trim_state == VDEV_TRIM_ACTIVE) {
4241 		(void) vdev_trim(vd, vd->vdev_trim_rate, vd->vdev_trim_partial,
4242 		    vd->vdev_trim_secure);
4243 	}
4244 	mutex_exit(&vd->vdev_trim_lock);
4245 
4246 	if (wasoffline ||
4247 	    (oldstate < VDEV_STATE_DEGRADED &&
4248 	    vd->vdev_state >= VDEV_STATE_DEGRADED)) {
4249 		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE);
4250 
4251 		/*
4252 		 * Asynchronously detach spare vdev if resilver or
4253 		 * rebuild is not required
4254 		 */
4255 		if (vd->vdev_unspare &&
4256 		    !dsl_scan_resilvering(spa->spa_dsl_pool) &&
4257 		    !dsl_scan_resilver_scheduled(spa->spa_dsl_pool) &&
4258 		    !vdev_rebuild_active(tvd))
4259 			spa_async_request(spa, SPA_ASYNC_DETACH_SPARE);
4260 	}
4261 	return (spa_vdev_state_exit(spa, vd, 0));
4262 }
4263 
4264 static int
4265 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
4266 {
4267 	vdev_t *vd, *tvd;
4268 	int error = 0;
4269 	uint64_t generation;
4270 	metaslab_group_t *mg;
4271 
4272 top:
4273 	spa_vdev_state_enter(spa, SCL_ALLOC);
4274 
4275 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
4276 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENODEV)));
4277 
4278 	if (!vd->vdev_ops->vdev_op_leaf)
4279 		return (spa_vdev_state_exit(spa, NULL, SET_ERROR(ENOTSUP)));
4280 
4281 	if (vd->vdev_ops == &vdev_draid_spare_ops)
4282 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
4283 
4284 	tvd = vd->vdev_top;
4285 	mg = tvd->vdev_mg;
4286 	generation = spa->spa_config_generation + 1;
4287 
4288 	/*
4289 	 * If the device isn't already offline, try to offline it.
4290 	 */
4291 	if (!vd->vdev_offline) {
4292 		/*
4293 		 * If this device has the only valid copy of some data,
4294 		 * don't allow it to be offlined. Log devices are always
4295 		 * expendable.
4296 		 */
4297 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
4298 		    vdev_dtl_required(vd))
4299 			return (spa_vdev_state_exit(spa, NULL,
4300 			    SET_ERROR(EBUSY)));
4301 
4302 		/*
4303 		 * If the top-level is a slog and it has had allocations
4304 		 * then proceed.  We check that the vdev's metaslab group
4305 		 * is not NULL since it's possible that we may have just
4306 		 * added this vdev but not yet initialized its metaslabs.
4307 		 */
4308 		if (tvd->vdev_islog && mg != NULL) {
4309 			/*
4310 			 * Prevent any future allocations.
4311 			 */
4312 			ASSERT3P(tvd->vdev_log_mg, ==, NULL);
4313 			metaslab_group_passivate(mg);
4314 			(void) spa_vdev_state_exit(spa, vd, 0);
4315 
4316 			error = spa_reset_logs(spa);
4317 
4318 			/*
4319 			 * If the log device was successfully reset but has
4320 			 * checkpointed data, do not offline it.
4321 			 */
4322 			if (error == 0 &&
4323 			    tvd->vdev_checkpoint_sm != NULL) {
4324 				ASSERT3U(space_map_allocated(
4325 				    tvd->vdev_checkpoint_sm), !=, 0);
4326 				error = ZFS_ERR_CHECKPOINT_EXISTS;
4327 			}
4328 
4329 			spa_vdev_state_enter(spa, SCL_ALLOC);
4330 
4331 			/*
4332 			 * Check to see if the config has changed.
4333 			 */
4334 			if (error || generation != spa->spa_config_generation) {
4335 				metaslab_group_activate(mg);
4336 				if (error)
4337 					return (spa_vdev_state_exit(spa,
4338 					    vd, error));
4339 				(void) spa_vdev_state_exit(spa, vd, 0);
4340 				goto top;
4341 			}
4342 			ASSERT0(tvd->vdev_stat.vs_alloc);
4343 		}
4344 
4345 		/*
4346 		 * Offline this device and reopen its top-level vdev.
4347 		 * If the top-level vdev is a log device then just offline
4348 		 * it. Otherwise, if this action results in the top-level
4349 		 * vdev becoming unusable, undo it and fail the request.
4350 		 */
4351 		vd->vdev_offline = B_TRUE;
4352 		vdev_reopen(tvd);
4353 
4354 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
4355 		    vdev_is_dead(tvd)) {
4356 			vd->vdev_offline = B_FALSE;
4357 			vdev_reopen(tvd);
4358 			return (spa_vdev_state_exit(spa, NULL,
4359 			    SET_ERROR(EBUSY)));
4360 		}
4361 
4362 		/*
4363 		 * Add the device back into the metaslab rotor so that
4364 		 * once we online the device it's open for business.
4365 		 */
4366 		if (tvd->vdev_islog && mg != NULL)
4367 			metaslab_group_activate(mg);
4368 	}
4369 
4370 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
4371 
4372 	return (spa_vdev_state_exit(spa, vd, 0));
4373 }
4374 
4375 int
4376 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
4377 {
4378 	int error;
4379 
4380 	mutex_enter(&spa->spa_vdev_top_lock);
4381 	error = vdev_offline_locked(spa, guid, flags);
4382 	mutex_exit(&spa->spa_vdev_top_lock);
4383 
4384 	return (error);
4385 }
4386 
4387 /*
4388  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
4389  * vdev_offline(), we assume the spa config is locked.  We also clear all
4390  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
4391  */
4392 void
4393 vdev_clear(spa_t *spa, vdev_t *vd)
4394 {
4395 	vdev_t *rvd = spa->spa_root_vdev;
4396 
4397 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
4398 
4399 	if (vd == NULL)
4400 		vd = rvd;
4401 
4402 	vd->vdev_stat.vs_read_errors = 0;
4403 	vd->vdev_stat.vs_write_errors = 0;
4404 	vd->vdev_stat.vs_checksum_errors = 0;
4405 	vd->vdev_stat.vs_slow_ios = 0;
4406 
4407 	for (int c = 0; c < vd->vdev_children; c++)
4408 		vdev_clear(spa, vd->vdev_child[c]);
4409 
4410 	/*
4411 	 * It makes no sense to "clear" an indirect  or removed vdev.
4412 	 */
4413 	if (!vdev_is_concrete(vd) || vd->vdev_removed)
4414 		return;
4415 
4416 	/*
4417 	 * If we're in the FAULTED state or have experienced failed I/O, then
4418 	 * clear the persistent state and attempt to reopen the device.  We
4419 	 * also mark the vdev config dirty, so that the new faulted state is
4420 	 * written out to disk.
4421 	 */
4422 	if (vd->vdev_faulted || vd->vdev_degraded ||
4423 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
4424 		/*
4425 		 * When reopening in response to a clear event, it may be due to
4426 		 * a fmadm repair request.  In this case, if the device is
4427 		 * still broken, we want to still post the ereport again.
4428 		 */
4429 		vd->vdev_forcefault = B_TRUE;
4430 
4431 		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
4432 		vd->vdev_cant_read = B_FALSE;
4433 		vd->vdev_cant_write = B_FALSE;
4434 		vd->vdev_stat.vs_aux = 0;
4435 
4436 		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
4437 
4438 		vd->vdev_forcefault = B_FALSE;
4439 
4440 		if (vd != rvd && vdev_writeable(vd->vdev_top))
4441 			vdev_state_dirty(vd->vdev_top);
4442 
4443 		/* If a resilver isn't required, check if vdevs can be culled */
4444 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd) &&
4445 		    !dsl_scan_resilvering(spa->spa_dsl_pool) &&
4446 		    !dsl_scan_resilver_scheduled(spa->spa_dsl_pool))
4447 			spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
4448 
4449 		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR);
4450 	}
4451 
4452 	/*
4453 	 * When clearing a FMA-diagnosed fault, we always want to
4454 	 * unspare the device, as we assume that the original spare was
4455 	 * done in response to the FMA fault.
4456 	 */
4457 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
4458 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
4459 	    vd->vdev_parent->vdev_child[0] == vd)
4460 		vd->vdev_unspare = B_TRUE;
4461 
4462 	/* Clear recent error events cache (i.e. duplicate events tracking) */
4463 	zfs_ereport_clear(spa, vd);
4464 }
4465 
4466 boolean_t
4467 vdev_is_dead(vdev_t *vd)
4468 {
4469 	/*
4470 	 * Holes and missing devices are always considered "dead".
4471 	 * This simplifies the code since we don't have to check for
4472 	 * these types of devices in the various code paths.
4473 	 * Instead we rely on the fact that we skip over dead devices
4474 	 * before issuing I/O to them.
4475 	 */
4476 	return (vd->vdev_state < VDEV_STATE_DEGRADED ||
4477 	    vd->vdev_ops == &vdev_hole_ops ||
4478 	    vd->vdev_ops == &vdev_missing_ops);
4479 }
4480 
4481 boolean_t
4482 vdev_readable(vdev_t *vd)
4483 {
4484 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
4485 }
4486 
4487 boolean_t
4488 vdev_writeable(vdev_t *vd)
4489 {
4490 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write &&
4491 	    vdev_is_concrete(vd));
4492 }
4493 
4494 boolean_t
4495 vdev_allocatable(vdev_t *vd)
4496 {
4497 	uint64_t state = vd->vdev_state;
4498 
4499 	/*
4500 	 * We currently allow allocations from vdevs which may be in the
4501 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
4502 	 * fails to reopen then we'll catch it later when we're holding
4503 	 * the proper locks.  Note that we have to get the vdev state
4504 	 * in a local variable because although it changes atomically,
4505 	 * we're asking two separate questions about it.
4506 	 */
4507 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
4508 	    !vd->vdev_cant_write && vdev_is_concrete(vd) &&
4509 	    vd->vdev_mg->mg_initialized);
4510 }
4511 
4512 boolean_t
4513 vdev_accessible(vdev_t *vd, zio_t *zio)
4514 {
4515 	ASSERT(zio->io_vd == vd);
4516 
4517 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
4518 		return (B_FALSE);
4519 
4520 	if (zio->io_type == ZIO_TYPE_READ)
4521 		return (!vd->vdev_cant_read);
4522 
4523 	if (zio->io_type == ZIO_TYPE_WRITE)
4524 		return (!vd->vdev_cant_write);
4525 
4526 	return (B_TRUE);
4527 }
4528 
4529 static void
4530 vdev_get_child_stat(vdev_t *cvd, vdev_stat_t *vs, vdev_stat_t *cvs)
4531 {
4532 	/*
4533 	 * Exclude the dRAID spare when aggregating to avoid double counting
4534 	 * the ops and bytes.  These IOs are counted by the physical leaves.
4535 	 */
4536 	if (cvd->vdev_ops == &vdev_draid_spare_ops)
4537 		return;
4538 
4539 	for (int t = 0; t < VS_ZIO_TYPES; t++) {
4540 		vs->vs_ops[t] += cvs->vs_ops[t];
4541 		vs->vs_bytes[t] += cvs->vs_bytes[t];
4542 	}
4543 
4544 	cvs->vs_scan_removing = cvd->vdev_removing;
4545 }
4546 
4547 /*
4548  * Get extended stats
4549  */
4550 static void
4551 vdev_get_child_stat_ex(vdev_t *cvd, vdev_stat_ex_t *vsx, vdev_stat_ex_t *cvsx)
4552 {
4553 	(void) cvd;
4554 
4555 	int t, b;
4556 	for (t = 0; t < ZIO_TYPES; t++) {
4557 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_disk_histo[0]); b++)
4558 			vsx->vsx_disk_histo[t][b] += cvsx->vsx_disk_histo[t][b];
4559 
4560 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_total_histo[0]); b++) {
4561 			vsx->vsx_total_histo[t][b] +=
4562 			    cvsx->vsx_total_histo[t][b];
4563 		}
4564 	}
4565 
4566 	for (t = 0; t < ZIO_PRIORITY_NUM_QUEUEABLE; t++) {
4567 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_queue_histo[0]); b++) {
4568 			vsx->vsx_queue_histo[t][b] +=
4569 			    cvsx->vsx_queue_histo[t][b];
4570 		}
4571 		vsx->vsx_active_queue[t] += cvsx->vsx_active_queue[t];
4572 		vsx->vsx_pend_queue[t] += cvsx->vsx_pend_queue[t];
4573 
4574 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_ind_histo[0]); b++)
4575 			vsx->vsx_ind_histo[t][b] += cvsx->vsx_ind_histo[t][b];
4576 
4577 		for (b = 0; b < ARRAY_SIZE(vsx->vsx_agg_histo[0]); b++)
4578 			vsx->vsx_agg_histo[t][b] += cvsx->vsx_agg_histo[t][b];
4579 	}
4580 
4581 }
4582 
4583 boolean_t
4584 vdev_is_spacemap_addressable(vdev_t *vd)
4585 {
4586 	if (spa_feature_is_active(vd->vdev_spa, SPA_FEATURE_SPACEMAP_V2))
4587 		return (B_TRUE);
4588 
4589 	/*
4590 	 * If double-word space map entries are not enabled we assume
4591 	 * 47 bits of the space map entry are dedicated to the entry's
4592 	 * offset (see SM_OFFSET_BITS in space_map.h). We then use that
4593 	 * to calculate the maximum address that can be described by a
4594 	 * space map entry for the given device.
4595 	 */
4596 	uint64_t shift = vd->vdev_ashift + SM_OFFSET_BITS;
4597 
4598 	if (shift >= 63) /* detect potential overflow */
4599 		return (B_TRUE);
4600 
4601 	return (vd->vdev_asize < (1ULL << shift));
4602 }
4603 
4604 /*
4605  * Get statistics for the given vdev.
4606  */
4607 static void
4608 vdev_get_stats_ex_impl(vdev_t *vd, vdev_stat_t *vs, vdev_stat_ex_t *vsx)
4609 {
4610 	int t;
4611 	/*
4612 	 * If we're getting stats on the root vdev, aggregate the I/O counts
4613 	 * over all top-level vdevs (i.e. the direct children of the root).
4614 	 */
4615 	if (!vd->vdev_ops->vdev_op_leaf) {
4616 		if (vs) {
4617 			memset(vs->vs_ops, 0, sizeof (vs->vs_ops));
4618 			memset(vs->vs_bytes, 0, sizeof (vs->vs_bytes));
4619 		}
4620 		if (vsx)
4621 			memset(vsx, 0, sizeof (*vsx));
4622 
4623 		for (int c = 0; c < vd->vdev_children; c++) {
4624 			vdev_t *cvd = vd->vdev_child[c];
4625 			vdev_stat_t *cvs = &cvd->vdev_stat;
4626 			vdev_stat_ex_t *cvsx = &cvd->vdev_stat_ex;
4627 
4628 			vdev_get_stats_ex_impl(cvd, cvs, cvsx);
4629 			if (vs)
4630 				vdev_get_child_stat(cvd, vs, cvs);
4631 			if (vsx)
4632 				vdev_get_child_stat_ex(cvd, vsx, cvsx);
4633 		}
4634 	} else {
4635 		/*
4636 		 * We're a leaf.  Just copy our ZIO active queue stats in.  The
4637 		 * other leaf stats are updated in vdev_stat_update().
4638 		 */
4639 		if (!vsx)
4640 			return;
4641 
4642 		memcpy(vsx, &vd->vdev_stat_ex, sizeof (vd->vdev_stat_ex));
4643 
4644 		for (t = 0; t < ZIO_PRIORITY_NUM_QUEUEABLE; t++) {
4645 			vsx->vsx_active_queue[t] = vd->vdev_queue.vq_cactive[t];
4646 			vsx->vsx_pend_queue[t] = vdev_queue_class_length(vd, t);
4647 		}
4648 	}
4649 }
4650 
4651 void
4652 vdev_get_stats_ex(vdev_t *vd, vdev_stat_t *vs, vdev_stat_ex_t *vsx)
4653 {
4654 	vdev_t *tvd = vd->vdev_top;
4655 	mutex_enter(&vd->vdev_stat_lock);
4656 	if (vs) {
4657 		memcpy(vs, &vd->vdev_stat, sizeof (*vs));
4658 		vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
4659 		vs->vs_state = vd->vdev_state;
4660 		vs->vs_rsize = vdev_get_min_asize(vd);
4661 
4662 		if (vd->vdev_ops->vdev_op_leaf) {
4663 			vs->vs_pspace = vd->vdev_psize;
4664 			vs->vs_rsize += VDEV_LABEL_START_SIZE +
4665 			    VDEV_LABEL_END_SIZE;
4666 			/*
4667 			 * Report initializing progress. Since we don't
4668 			 * have the initializing locks held, this is only
4669 			 * an estimate (although a fairly accurate one).
4670 			 */
4671 			vs->vs_initialize_bytes_done =
4672 			    vd->vdev_initialize_bytes_done;
4673 			vs->vs_initialize_bytes_est =
4674 			    vd->vdev_initialize_bytes_est;
4675 			vs->vs_initialize_state = vd->vdev_initialize_state;
4676 			vs->vs_initialize_action_time =
4677 			    vd->vdev_initialize_action_time;
4678 
4679 			/*
4680 			 * Report manual TRIM progress. Since we don't have
4681 			 * the manual TRIM locks held, this is only an
4682 			 * estimate (although fairly accurate one).
4683 			 */
4684 			vs->vs_trim_notsup = !vd->vdev_has_trim;
4685 			vs->vs_trim_bytes_done = vd->vdev_trim_bytes_done;
4686 			vs->vs_trim_bytes_est = vd->vdev_trim_bytes_est;
4687 			vs->vs_trim_state = vd->vdev_trim_state;
4688 			vs->vs_trim_action_time = vd->vdev_trim_action_time;
4689 
4690 			/* Set when there is a deferred resilver. */
4691 			vs->vs_resilver_deferred = vd->vdev_resilver_deferred;
4692 		}
4693 
4694 		/*
4695 		 * Report expandable space on top-level, non-auxiliary devices
4696 		 * only. The expandable space is reported in terms of metaslab
4697 		 * sized units since that determines how much space the pool
4698 		 * can expand.
4699 		 */
4700 		if (vd->vdev_aux == NULL && tvd != NULL) {
4701 			vs->vs_esize = P2ALIGN(
4702 			    vd->vdev_max_asize - vd->vdev_asize,
4703 			    1ULL << tvd->vdev_ms_shift);
4704 		}
4705 
4706 		vs->vs_configured_ashift = vd->vdev_top != NULL
4707 		    ? vd->vdev_top->vdev_ashift : vd->vdev_ashift;
4708 		vs->vs_logical_ashift = vd->vdev_logical_ashift;
4709 		if (vd->vdev_physical_ashift <= ASHIFT_MAX)
4710 			vs->vs_physical_ashift = vd->vdev_physical_ashift;
4711 		else
4712 			vs->vs_physical_ashift = 0;
4713 
4714 		/*
4715 		 * Report fragmentation and rebuild progress for top-level,
4716 		 * non-auxiliary, concrete devices.
4717 		 */
4718 		if (vd->vdev_aux == NULL && vd == vd->vdev_top &&
4719 		    vdev_is_concrete(vd)) {
4720 			/*
4721 			 * The vdev fragmentation rating doesn't take into
4722 			 * account the embedded slog metaslab (vdev_log_mg).
4723 			 * Since it's only one metaslab, it would have a tiny
4724 			 * impact on the overall fragmentation.
4725 			 */
4726 			vs->vs_fragmentation = (vd->vdev_mg != NULL) ?
4727 			    vd->vdev_mg->mg_fragmentation : 0;
4728 		}
4729 		vs->vs_noalloc = MAX(vd->vdev_noalloc,
4730 		    tvd ? tvd->vdev_noalloc : 0);
4731 	}
4732 
4733 	vdev_get_stats_ex_impl(vd, vs, vsx);
4734 	mutex_exit(&vd->vdev_stat_lock);
4735 }
4736 
4737 void
4738 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
4739 {
4740 	return (vdev_get_stats_ex(vd, vs, NULL));
4741 }
4742 
4743 void
4744 vdev_clear_stats(vdev_t *vd)
4745 {
4746 	mutex_enter(&vd->vdev_stat_lock);
4747 	vd->vdev_stat.vs_space = 0;
4748 	vd->vdev_stat.vs_dspace = 0;
4749 	vd->vdev_stat.vs_alloc = 0;
4750 	mutex_exit(&vd->vdev_stat_lock);
4751 }
4752 
4753 void
4754 vdev_scan_stat_init(vdev_t *vd)
4755 {
4756 	vdev_stat_t *vs = &vd->vdev_stat;
4757 
4758 	for (int c = 0; c < vd->vdev_children; c++)
4759 		vdev_scan_stat_init(vd->vdev_child[c]);
4760 
4761 	mutex_enter(&vd->vdev_stat_lock);
4762 	vs->vs_scan_processed = 0;
4763 	mutex_exit(&vd->vdev_stat_lock);
4764 }
4765 
4766 void
4767 vdev_stat_update(zio_t *zio, uint64_t psize)
4768 {
4769 	spa_t *spa = zio->io_spa;
4770 	vdev_t *rvd = spa->spa_root_vdev;
4771 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
4772 	vdev_t *pvd;
4773 	uint64_t txg = zio->io_txg;
4774 /* Suppress ASAN false positive */
4775 #ifdef __SANITIZE_ADDRESS__
4776 	vdev_stat_t *vs = vd ? &vd->vdev_stat : NULL;
4777 	vdev_stat_ex_t *vsx = vd ? &vd->vdev_stat_ex : NULL;
4778 #else
4779 	vdev_stat_t *vs = &vd->vdev_stat;
4780 	vdev_stat_ex_t *vsx = &vd->vdev_stat_ex;
4781 #endif
4782 	zio_type_t type = zio->io_type;
4783 	int flags = zio->io_flags;
4784 
4785 	/*
4786 	 * If this i/o is a gang leader, it didn't do any actual work.
4787 	 */
4788 	if (zio->io_gang_tree)
4789 		return;
4790 
4791 	if (zio->io_error == 0) {
4792 		/*
4793 		 * If this is a root i/o, don't count it -- we've already
4794 		 * counted the top-level vdevs, and vdev_get_stats() will
4795 		 * aggregate them when asked.  This reduces contention on
4796 		 * the root vdev_stat_lock and implicitly handles blocks
4797 		 * that compress away to holes, for which there is no i/o.
4798 		 * (Holes never create vdev children, so all the counters
4799 		 * remain zero, which is what we want.)
4800 		 *
4801 		 * Note: this only applies to successful i/o (io_error == 0)
4802 		 * because unlike i/o counts, errors are not additive.
4803 		 * When reading a ditto block, for example, failure of
4804 		 * one top-level vdev does not imply a root-level error.
4805 		 */
4806 		if (vd == rvd)
4807 			return;
4808 
4809 		ASSERT(vd == zio->io_vd);
4810 
4811 		if (flags & ZIO_FLAG_IO_BYPASS)
4812 			return;
4813 
4814 		mutex_enter(&vd->vdev_stat_lock);
4815 
4816 		if (flags & ZIO_FLAG_IO_REPAIR) {
4817 			/*
4818 			 * Repair is the result of a resilver issued by the
4819 			 * scan thread (spa_sync).
4820 			 */
4821 			if (flags & ZIO_FLAG_SCAN_THREAD) {
4822 				dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
4823 				dsl_scan_phys_t *scn_phys = &scn->scn_phys;
4824 				uint64_t *processed = &scn_phys->scn_processed;
4825 
4826 				if (vd->vdev_ops->vdev_op_leaf)
4827 					atomic_add_64(processed, psize);
4828 				vs->vs_scan_processed += psize;
4829 			}
4830 
4831 			/*
4832 			 * Repair is the result of a rebuild issued by the
4833 			 * rebuild thread (vdev_rebuild_thread).  To avoid
4834 			 * double counting repaired bytes the virtual dRAID
4835 			 * spare vdev is excluded from the processed bytes.
4836 			 */
4837 			if (zio->io_priority == ZIO_PRIORITY_REBUILD) {
4838 				vdev_t *tvd = vd->vdev_top;
4839 				vdev_rebuild_t *vr = &tvd->vdev_rebuild_config;
4840 				vdev_rebuild_phys_t *vrp = &vr->vr_rebuild_phys;
4841 				uint64_t *rebuilt = &vrp->vrp_bytes_rebuilt;
4842 
4843 				if (vd->vdev_ops->vdev_op_leaf &&
4844 				    vd->vdev_ops != &vdev_draid_spare_ops) {
4845 					atomic_add_64(rebuilt, psize);
4846 				}
4847 				vs->vs_rebuild_processed += psize;
4848 			}
4849 
4850 			if (flags & ZIO_FLAG_SELF_HEAL)
4851 				vs->vs_self_healed += psize;
4852 		}
4853 
4854 		/*
4855 		 * The bytes/ops/histograms are recorded at the leaf level and
4856 		 * aggregated into the higher level vdevs in vdev_get_stats().
4857 		 */
4858 		if (vd->vdev_ops->vdev_op_leaf &&
4859 		    (zio->io_priority < ZIO_PRIORITY_NUM_QUEUEABLE)) {
4860 			zio_type_t vs_type = type;
4861 			zio_priority_t priority = zio->io_priority;
4862 
4863 			/*
4864 			 * TRIM ops and bytes are reported to user space as
4865 			 * ZIO_TYPE_IOCTL.  This is done to preserve the
4866 			 * vdev_stat_t structure layout for user space.
4867 			 */
4868 			if (type == ZIO_TYPE_TRIM)
4869 				vs_type = ZIO_TYPE_IOCTL;
4870 
4871 			/*
4872 			 * Solely for the purposes of 'zpool iostat -lqrw'
4873 			 * reporting use the priority to categorize the IO.
4874 			 * Only the following are reported to user space:
4875 			 *
4876 			 *   ZIO_PRIORITY_SYNC_READ,
4877 			 *   ZIO_PRIORITY_SYNC_WRITE,
4878 			 *   ZIO_PRIORITY_ASYNC_READ,
4879 			 *   ZIO_PRIORITY_ASYNC_WRITE,
4880 			 *   ZIO_PRIORITY_SCRUB,
4881 			 *   ZIO_PRIORITY_TRIM,
4882 			 *   ZIO_PRIORITY_REBUILD.
4883 			 */
4884 			if (priority == ZIO_PRIORITY_INITIALIZING) {
4885 				ASSERT3U(type, ==, ZIO_TYPE_WRITE);
4886 				priority = ZIO_PRIORITY_ASYNC_WRITE;
4887 			} else if (priority == ZIO_PRIORITY_REMOVAL) {
4888 				priority = ((type == ZIO_TYPE_WRITE) ?
4889 				    ZIO_PRIORITY_ASYNC_WRITE :
4890 				    ZIO_PRIORITY_ASYNC_READ);
4891 			}
4892 
4893 			vs->vs_ops[vs_type]++;
4894 			vs->vs_bytes[vs_type] += psize;
4895 
4896 			if (flags & ZIO_FLAG_DELEGATED) {
4897 				vsx->vsx_agg_histo[priority]
4898 				    [RQ_HISTO(zio->io_size)]++;
4899 			} else {
4900 				vsx->vsx_ind_histo[priority]
4901 				    [RQ_HISTO(zio->io_size)]++;
4902 			}
4903 
4904 			if (zio->io_delta && zio->io_delay) {
4905 				vsx->vsx_queue_histo[priority]
4906 				    [L_HISTO(zio->io_delta - zio->io_delay)]++;
4907 				vsx->vsx_disk_histo[type]
4908 				    [L_HISTO(zio->io_delay)]++;
4909 				vsx->vsx_total_histo[type]
4910 				    [L_HISTO(zio->io_delta)]++;
4911 			}
4912 		}
4913 
4914 		mutex_exit(&vd->vdev_stat_lock);
4915 		return;
4916 	}
4917 
4918 	if (flags & ZIO_FLAG_SPECULATIVE)
4919 		return;
4920 
4921 	/*
4922 	 * If this is an I/O error that is going to be retried, then ignore the
4923 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
4924 	 * hard errors, when in reality they can happen for any number of
4925 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
4926 	 */
4927 	if (zio->io_error == EIO &&
4928 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
4929 		return;
4930 
4931 	/*
4932 	 * Intent logs writes won't propagate their error to the root
4933 	 * I/O so don't mark these types of failures as pool-level
4934 	 * errors.
4935 	 */
4936 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
4937 		return;
4938 
4939 	if (type == ZIO_TYPE_WRITE && txg != 0 &&
4940 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
4941 	    (flags & ZIO_FLAG_SCAN_THREAD) ||
4942 	    spa->spa_claiming)) {
4943 		/*
4944 		 * This is either a normal write (not a repair), or it's
4945 		 * a repair induced by the scrub thread, or it's a repair
4946 		 * made by zil_claim() during spa_load() in the first txg.
4947 		 * In the normal case, we commit the DTL change in the same
4948 		 * txg as the block was born.  In the scrub-induced repair
4949 		 * case, we know that scrubs run in first-pass syncing context,
4950 		 * so we commit the DTL change in spa_syncing_txg(spa).
4951 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
4952 		 *
4953 		 * We currently do not make DTL entries for failed spontaneous
4954 		 * self-healing writes triggered by normal (non-scrubbing)
4955 		 * reads, because we have no transactional context in which to
4956 		 * do so -- and it's not clear that it'd be desirable anyway.
4957 		 */
4958 		if (vd->vdev_ops->vdev_op_leaf) {
4959 			uint64_t commit_txg = txg;
4960 			if (flags & ZIO_FLAG_SCAN_THREAD) {
4961 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
4962 				ASSERT(spa_sync_pass(spa) == 1);
4963 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
4964 				commit_txg = spa_syncing_txg(spa);
4965 			} else if (spa->spa_claiming) {
4966 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
4967 				commit_txg = spa_first_txg(spa);
4968 			}
4969 			ASSERT(commit_txg >= spa_syncing_txg(spa));
4970 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
4971 				return;
4972 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
4973 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
4974 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
4975 		}
4976 		if (vd != rvd)
4977 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
4978 	}
4979 }
4980 
4981 int64_t
4982 vdev_deflated_space(vdev_t *vd, int64_t space)
4983 {
4984 	ASSERT((space & (SPA_MINBLOCKSIZE-1)) == 0);
4985 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
4986 
4987 	return ((space >> SPA_MINBLOCKSHIFT) * vd->vdev_deflate_ratio);
4988 }
4989 
4990 /*
4991  * Update the in-core space usage stats for this vdev, its metaslab class,
4992  * and the root vdev.
4993  */
4994 void
4995 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
4996     int64_t space_delta)
4997 {
4998 	(void) defer_delta;
4999 	int64_t dspace_delta;
5000 	spa_t *spa = vd->vdev_spa;
5001 	vdev_t *rvd = spa->spa_root_vdev;
5002 
5003 	ASSERT(vd == vd->vdev_top);
5004 
5005 	/*
5006 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
5007 	 * factor.  We must calculate this here and not at the root vdev
5008 	 * because the root vdev's psize-to-asize is simply the max of its
5009 	 * children's, thus not accurate enough for us.
5010 	 */
5011 	dspace_delta = vdev_deflated_space(vd, space_delta);
5012 
5013 	mutex_enter(&vd->vdev_stat_lock);
5014 	/* ensure we won't underflow */
5015 	if (alloc_delta < 0) {
5016 		ASSERT3U(vd->vdev_stat.vs_alloc, >=, -alloc_delta);
5017 	}
5018 
5019 	vd->vdev_stat.vs_alloc += alloc_delta;
5020 	vd->vdev_stat.vs_space += space_delta;
5021 	vd->vdev_stat.vs_dspace += dspace_delta;
5022 	mutex_exit(&vd->vdev_stat_lock);
5023 
5024 	/* every class but log contributes to root space stats */
5025 	if (vd->vdev_mg != NULL && !vd->vdev_islog) {
5026 		ASSERT(!vd->vdev_isl2cache);
5027 		mutex_enter(&rvd->vdev_stat_lock);
5028 		rvd->vdev_stat.vs_alloc += alloc_delta;
5029 		rvd->vdev_stat.vs_space += space_delta;
5030 		rvd->vdev_stat.vs_dspace += dspace_delta;
5031 		mutex_exit(&rvd->vdev_stat_lock);
5032 	}
5033 	/* Note: metaslab_class_space_update moved to metaslab_space_update */
5034 }
5035 
5036 /*
5037  * Mark a top-level vdev's config as dirty, placing it on the dirty list
5038  * so that it will be written out next time the vdev configuration is synced.
5039  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
5040  */
5041 void
5042 vdev_config_dirty(vdev_t *vd)
5043 {
5044 	spa_t *spa = vd->vdev_spa;
5045 	vdev_t *rvd = spa->spa_root_vdev;
5046 	int c;
5047 
5048 	ASSERT(spa_writeable(spa));
5049 
5050 	/*
5051 	 * If this is an aux vdev (as with l2cache and spare devices), then we
5052 	 * update the vdev config manually and set the sync flag.
5053 	 */
5054 	if (vd->vdev_aux != NULL) {
5055 		spa_aux_vdev_t *sav = vd->vdev_aux;
5056 		nvlist_t **aux;
5057 		uint_t naux;
5058 
5059 		for (c = 0; c < sav->sav_count; c++) {
5060 			if (sav->sav_vdevs[c] == vd)
5061 				break;
5062 		}
5063 
5064 		if (c == sav->sav_count) {
5065 			/*
5066 			 * We're being removed.  There's nothing more to do.
5067 			 */
5068 			ASSERT(sav->sav_sync == B_TRUE);
5069 			return;
5070 		}
5071 
5072 		sav->sav_sync = B_TRUE;
5073 
5074 		if (nvlist_lookup_nvlist_array(sav->sav_config,
5075 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
5076 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
5077 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
5078 		}
5079 
5080 		ASSERT(c < naux);
5081 
5082 		/*
5083 		 * Setting the nvlist in the middle if the array is a little
5084 		 * sketchy, but it will work.
5085 		 */
5086 		nvlist_free(aux[c]);
5087 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
5088 
5089 		return;
5090 	}
5091 
5092 	/*
5093 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
5094 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
5095 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
5096 	 * so this is sufficient to ensure mutual exclusion.
5097 	 */
5098 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
5099 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
5100 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
5101 
5102 	if (vd == rvd) {
5103 		for (c = 0; c < rvd->vdev_children; c++)
5104 			vdev_config_dirty(rvd->vdev_child[c]);
5105 	} else {
5106 		ASSERT(vd == vd->vdev_top);
5107 
5108 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
5109 		    vdev_is_concrete(vd)) {
5110 			list_insert_head(&spa->spa_config_dirty_list, vd);
5111 		}
5112 	}
5113 }
5114 
5115 void
5116 vdev_config_clean(vdev_t *vd)
5117 {
5118 	spa_t *spa = vd->vdev_spa;
5119 
5120 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
5121 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
5122 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
5123 
5124 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
5125 	list_remove(&spa->spa_config_dirty_list, vd);
5126 }
5127 
5128 /*
5129  * Mark a top-level vdev's state as dirty, so that the next pass of
5130  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
5131  * the state changes from larger config changes because they require
5132  * much less locking, and are often needed for administrative actions.
5133  */
5134 void
5135 vdev_state_dirty(vdev_t *vd)
5136 {
5137 	spa_t *spa = vd->vdev_spa;
5138 
5139 	ASSERT(spa_writeable(spa));
5140 	ASSERT(vd == vd->vdev_top);
5141 
5142 	/*
5143 	 * The state list is protected by the SCL_STATE lock.  The caller
5144 	 * must either hold SCL_STATE as writer, or must be the sync thread
5145 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
5146 	 * so this is sufficient to ensure mutual exclusion.
5147 	 */
5148 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
5149 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
5150 	    spa_config_held(spa, SCL_STATE, RW_READER)));
5151 
5152 	if (!list_link_active(&vd->vdev_state_dirty_node) &&
5153 	    vdev_is_concrete(vd))
5154 		list_insert_head(&spa->spa_state_dirty_list, vd);
5155 }
5156 
5157 void
5158 vdev_state_clean(vdev_t *vd)
5159 {
5160 	spa_t *spa = vd->vdev_spa;
5161 
5162 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
5163 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
5164 	    spa_config_held(spa, SCL_STATE, RW_READER)));
5165 
5166 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
5167 	list_remove(&spa->spa_state_dirty_list, vd);
5168 }
5169 
5170 /*
5171  * Propagate vdev state up from children to parent.
5172  */
5173 void
5174 vdev_propagate_state(vdev_t *vd)
5175 {
5176 	spa_t *spa = vd->vdev_spa;
5177 	vdev_t *rvd = spa->spa_root_vdev;
5178 	int degraded = 0, faulted = 0;
5179 	int corrupted = 0;
5180 	vdev_t *child;
5181 
5182 	if (vd->vdev_children > 0) {
5183 		for (int c = 0; c < vd->vdev_children; c++) {
5184 			child = vd->vdev_child[c];
5185 
5186 			/*
5187 			 * Don't factor holes or indirect vdevs into the
5188 			 * decision.
5189 			 */
5190 			if (!vdev_is_concrete(child))
5191 				continue;
5192 
5193 			if (!vdev_readable(child) ||
5194 			    (!vdev_writeable(child) && spa_writeable(spa))) {
5195 				/*
5196 				 * Root special: if there is a top-level log
5197 				 * device, treat the root vdev as if it were
5198 				 * degraded.
5199 				 */
5200 				if (child->vdev_islog && vd == rvd)
5201 					degraded++;
5202 				else
5203 					faulted++;
5204 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
5205 				degraded++;
5206 			}
5207 
5208 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
5209 				corrupted++;
5210 		}
5211 
5212 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
5213 
5214 		/*
5215 		 * Root special: if there is a top-level vdev that cannot be
5216 		 * opened due to corrupted metadata, then propagate the root
5217 		 * vdev's aux state as 'corrupt' rather than 'insufficient
5218 		 * replicas'.
5219 		 */
5220 		if (corrupted && vd == rvd &&
5221 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
5222 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
5223 			    VDEV_AUX_CORRUPT_DATA);
5224 	}
5225 
5226 	if (vd->vdev_parent)
5227 		vdev_propagate_state(vd->vdev_parent);
5228 }
5229 
5230 /*
5231  * Set a vdev's state.  If this is during an open, we don't update the parent
5232  * state, because we're in the process of opening children depth-first.
5233  * Otherwise, we propagate the change to the parent.
5234  *
5235  * If this routine places a device in a faulted state, an appropriate ereport is
5236  * generated.
5237  */
5238 void
5239 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
5240 {
5241 	uint64_t save_state;
5242 	spa_t *spa = vd->vdev_spa;
5243 
5244 	if (state == vd->vdev_state) {
5245 		/*
5246 		 * Since vdev_offline() code path is already in an offline
5247 		 * state we can miss a statechange event to OFFLINE. Check
5248 		 * the previous state to catch this condition.
5249 		 */
5250 		if (vd->vdev_ops->vdev_op_leaf &&
5251 		    (state == VDEV_STATE_OFFLINE) &&
5252 		    (vd->vdev_prevstate >= VDEV_STATE_FAULTED)) {
5253 			/* post an offline state change */
5254 			zfs_post_state_change(spa, vd, vd->vdev_prevstate);
5255 		}
5256 		vd->vdev_stat.vs_aux = aux;
5257 		return;
5258 	}
5259 
5260 	save_state = vd->vdev_state;
5261 
5262 	vd->vdev_state = state;
5263 	vd->vdev_stat.vs_aux = aux;
5264 
5265 	/*
5266 	 * If we are setting the vdev state to anything but an open state, then
5267 	 * always close the underlying device unless the device has requested
5268 	 * a delayed close (i.e. we're about to remove or fault the device).
5269 	 * Otherwise, we keep accessible but invalid devices open forever.
5270 	 * We don't call vdev_close() itself, because that implies some extra
5271 	 * checks (offline, etc) that we don't want here.  This is limited to
5272 	 * leaf devices, because otherwise closing the device will affect other
5273 	 * children.
5274 	 */
5275 	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
5276 	    vd->vdev_ops->vdev_op_leaf)
5277 		vd->vdev_ops->vdev_op_close(vd);
5278 
5279 	if (vd->vdev_removed &&
5280 	    state == VDEV_STATE_CANT_OPEN &&
5281 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
5282 		/*
5283 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
5284 		 * device was previously marked removed and someone attempted to
5285 		 * reopen it.  If this failed due to a nonexistent device, then
5286 		 * keep the device in the REMOVED state.  We also let this be if
5287 		 * it is one of our special test online cases, which is only
5288 		 * attempting to online the device and shouldn't generate an FMA
5289 		 * fault.
5290 		 */
5291 		vd->vdev_state = VDEV_STATE_REMOVED;
5292 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
5293 	} else if (state == VDEV_STATE_REMOVED) {
5294 		vd->vdev_removed = B_TRUE;
5295 	} else if (state == VDEV_STATE_CANT_OPEN) {
5296 		/*
5297 		 * If we fail to open a vdev during an import or recovery, we
5298 		 * mark it as "not available", which signifies that it was
5299 		 * never there to begin with.  Failure to open such a device
5300 		 * is not considered an error.
5301 		 */
5302 		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
5303 		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
5304 		    vd->vdev_ops->vdev_op_leaf)
5305 			vd->vdev_not_present = 1;
5306 
5307 		/*
5308 		 * Post the appropriate ereport.  If the 'prevstate' field is
5309 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
5310 		 * that this is part of a vdev_reopen().  In this case, we don't
5311 		 * want to post the ereport if the device was already in the
5312 		 * CANT_OPEN state beforehand.
5313 		 *
5314 		 * If the 'checkremove' flag is set, then this is an attempt to
5315 		 * online the device in response to an insertion event.  If we
5316 		 * hit this case, then we have detected an insertion event for a
5317 		 * faulted or offline device that wasn't in the removed state.
5318 		 * In this scenario, we don't post an ereport because we are
5319 		 * about to replace the device, or attempt an online with
5320 		 * vdev_forcefault, which will generate the fault for us.
5321 		 */
5322 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
5323 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
5324 		    vd != spa->spa_root_vdev) {
5325 			const char *class;
5326 
5327 			switch (aux) {
5328 			case VDEV_AUX_OPEN_FAILED:
5329 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
5330 				break;
5331 			case VDEV_AUX_CORRUPT_DATA:
5332 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
5333 				break;
5334 			case VDEV_AUX_NO_REPLICAS:
5335 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
5336 				break;
5337 			case VDEV_AUX_BAD_GUID_SUM:
5338 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
5339 				break;
5340 			case VDEV_AUX_TOO_SMALL:
5341 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
5342 				break;
5343 			case VDEV_AUX_BAD_LABEL:
5344 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
5345 				break;
5346 			case VDEV_AUX_BAD_ASHIFT:
5347 				class = FM_EREPORT_ZFS_DEVICE_BAD_ASHIFT;
5348 				break;
5349 			default:
5350 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
5351 			}
5352 
5353 			(void) zfs_ereport_post(class, spa, vd, NULL, NULL,
5354 			    save_state);
5355 		}
5356 
5357 		/* Erase any notion of persistent removed state */
5358 		vd->vdev_removed = B_FALSE;
5359 	} else {
5360 		vd->vdev_removed = B_FALSE;
5361 	}
5362 
5363 	/*
5364 	 * Notify ZED of any significant state-change on a leaf vdev.
5365 	 *
5366 	 */
5367 	if (vd->vdev_ops->vdev_op_leaf) {
5368 		/* preserve original state from a vdev_reopen() */
5369 		if ((vd->vdev_prevstate != VDEV_STATE_UNKNOWN) &&
5370 		    (vd->vdev_prevstate != vd->vdev_state) &&
5371 		    (save_state <= VDEV_STATE_CLOSED))
5372 			save_state = vd->vdev_prevstate;
5373 
5374 		/* filter out state change due to initial vdev_open */
5375 		if (save_state > VDEV_STATE_CLOSED)
5376 			zfs_post_state_change(spa, vd, save_state);
5377 	}
5378 
5379 	if (!isopen && vd->vdev_parent)
5380 		vdev_propagate_state(vd->vdev_parent);
5381 }
5382 
5383 boolean_t
5384 vdev_children_are_offline(vdev_t *vd)
5385 {
5386 	ASSERT(!vd->vdev_ops->vdev_op_leaf);
5387 
5388 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
5389 		if (vd->vdev_child[i]->vdev_state != VDEV_STATE_OFFLINE)
5390 			return (B_FALSE);
5391 	}
5392 
5393 	return (B_TRUE);
5394 }
5395 
5396 /*
5397  * Check the vdev configuration to ensure that it's capable of supporting
5398  * a root pool. We do not support partial configuration.
5399  */
5400 boolean_t
5401 vdev_is_bootable(vdev_t *vd)
5402 {
5403 	if (!vd->vdev_ops->vdev_op_leaf) {
5404 		const char *vdev_type = vd->vdev_ops->vdev_op_type;
5405 
5406 		if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0)
5407 			return (B_FALSE);
5408 	}
5409 
5410 	for (int c = 0; c < vd->vdev_children; c++) {
5411 		if (!vdev_is_bootable(vd->vdev_child[c]))
5412 			return (B_FALSE);
5413 	}
5414 	return (B_TRUE);
5415 }
5416 
5417 boolean_t
5418 vdev_is_concrete(vdev_t *vd)
5419 {
5420 	vdev_ops_t *ops = vd->vdev_ops;
5421 	if (ops == &vdev_indirect_ops || ops == &vdev_hole_ops ||
5422 	    ops == &vdev_missing_ops || ops == &vdev_root_ops) {
5423 		return (B_FALSE);
5424 	} else {
5425 		return (B_TRUE);
5426 	}
5427 }
5428 
5429 /*
5430  * Determine if a log device has valid content.  If the vdev was
5431  * removed or faulted in the MOS config then we know that
5432  * the content on the log device has already been written to the pool.
5433  */
5434 boolean_t
5435 vdev_log_state_valid(vdev_t *vd)
5436 {
5437 	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
5438 	    !vd->vdev_removed)
5439 		return (B_TRUE);
5440 
5441 	for (int c = 0; c < vd->vdev_children; c++)
5442 		if (vdev_log_state_valid(vd->vdev_child[c]))
5443 			return (B_TRUE);
5444 
5445 	return (B_FALSE);
5446 }
5447 
5448 /*
5449  * Expand a vdev if possible.
5450  */
5451 void
5452 vdev_expand(vdev_t *vd, uint64_t txg)
5453 {
5454 	ASSERT(vd->vdev_top == vd);
5455 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
5456 	ASSERT(vdev_is_concrete(vd));
5457 
5458 	vdev_set_deflate_ratio(vd);
5459 
5460 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count &&
5461 	    vdev_is_concrete(vd)) {
5462 		vdev_metaslab_group_create(vd);
5463 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
5464 		vdev_config_dirty(vd);
5465 	}
5466 }
5467 
5468 /*
5469  * Split a vdev.
5470  */
5471 void
5472 vdev_split(vdev_t *vd)
5473 {
5474 	vdev_t *cvd, *pvd = vd->vdev_parent;
5475 
5476 	VERIFY3U(pvd->vdev_children, >, 1);
5477 
5478 	vdev_remove_child(pvd, vd);
5479 	vdev_compact_children(pvd);
5480 
5481 	ASSERT3P(pvd->vdev_child, !=, NULL);
5482 
5483 	cvd = pvd->vdev_child[0];
5484 	if (pvd->vdev_children == 1) {
5485 		vdev_remove_parent(cvd);
5486 		cvd->vdev_splitting = B_TRUE;
5487 	}
5488 	vdev_propagate_state(cvd);
5489 }
5490 
5491 void
5492 vdev_deadman(vdev_t *vd, const char *tag)
5493 {
5494 	for (int c = 0; c < vd->vdev_children; c++) {
5495 		vdev_t *cvd = vd->vdev_child[c];
5496 
5497 		vdev_deadman(cvd, tag);
5498 	}
5499 
5500 	if (vd->vdev_ops->vdev_op_leaf) {
5501 		vdev_queue_t *vq = &vd->vdev_queue;
5502 
5503 		mutex_enter(&vq->vq_lock);
5504 		if (vq->vq_active > 0) {
5505 			spa_t *spa = vd->vdev_spa;
5506 			zio_t *fio;
5507 			uint64_t delta;
5508 
5509 			zfs_dbgmsg("slow vdev: %s has %u active IOs",
5510 			    vd->vdev_path, vq->vq_active);
5511 
5512 			/*
5513 			 * Look at the head of all the pending queues,
5514 			 * if any I/O has been outstanding for longer than
5515 			 * the spa_deadman_synctime invoke the deadman logic.
5516 			 */
5517 			fio = list_head(&vq->vq_active_list);
5518 			delta = gethrtime() - fio->io_timestamp;
5519 			if (delta > spa_deadman_synctime(spa))
5520 				zio_deadman(fio, tag);
5521 		}
5522 		mutex_exit(&vq->vq_lock);
5523 	}
5524 }
5525 
5526 void
5527 vdev_defer_resilver(vdev_t *vd)
5528 {
5529 	ASSERT(vd->vdev_ops->vdev_op_leaf);
5530 
5531 	vd->vdev_resilver_deferred = B_TRUE;
5532 	vd->vdev_spa->spa_resilver_deferred = B_TRUE;
5533 }
5534 
5535 /*
5536  * Clears the resilver deferred flag on all leaf devs under vd. Returns
5537  * B_TRUE if we have devices that need to be resilvered and are available to
5538  * accept resilver I/Os.
5539  */
5540 boolean_t
5541 vdev_clear_resilver_deferred(vdev_t *vd, dmu_tx_t *tx)
5542 {
5543 	boolean_t resilver_needed = B_FALSE;
5544 	spa_t *spa = vd->vdev_spa;
5545 
5546 	for (int c = 0; c < vd->vdev_children; c++) {
5547 		vdev_t *cvd = vd->vdev_child[c];
5548 		resilver_needed |= vdev_clear_resilver_deferred(cvd, tx);
5549 	}
5550 
5551 	if (vd == spa->spa_root_vdev &&
5552 	    spa_feature_is_active(spa, SPA_FEATURE_RESILVER_DEFER)) {
5553 		spa_feature_decr(spa, SPA_FEATURE_RESILVER_DEFER, tx);
5554 		vdev_config_dirty(vd);
5555 		spa->spa_resilver_deferred = B_FALSE;
5556 		return (resilver_needed);
5557 	}
5558 
5559 	if (!vdev_is_concrete(vd) || vd->vdev_aux ||
5560 	    !vd->vdev_ops->vdev_op_leaf)
5561 		return (resilver_needed);
5562 
5563 	vd->vdev_resilver_deferred = B_FALSE;
5564 
5565 	return (!vdev_is_dead(vd) && !vd->vdev_offline &&
5566 	    vdev_resilver_needed(vd, NULL, NULL));
5567 }
5568 
5569 boolean_t
5570 vdev_xlate_is_empty(range_seg64_t *rs)
5571 {
5572 	return (rs->rs_start == rs->rs_end);
5573 }
5574 
5575 /*
5576  * Translate a logical range to the first contiguous physical range for the
5577  * specified vdev_t.  This function is initially called with a leaf vdev and
5578  * will walk each parent vdev until it reaches a top-level vdev. Once the
5579  * top-level is reached the physical range is initialized and the recursive
5580  * function begins to unwind. As it unwinds it calls the parent's vdev
5581  * specific translation function to do the real conversion.
5582  */
5583 void
5584 vdev_xlate(vdev_t *vd, const range_seg64_t *logical_rs,
5585     range_seg64_t *physical_rs, range_seg64_t *remain_rs)
5586 {
5587 	/*
5588 	 * Walk up the vdev tree
5589 	 */
5590 	if (vd != vd->vdev_top) {
5591 		vdev_xlate(vd->vdev_parent, logical_rs, physical_rs,
5592 		    remain_rs);
5593 	} else {
5594 		/*
5595 		 * We've reached the top-level vdev, initialize the physical
5596 		 * range to the logical range and set an empty remaining
5597 		 * range then start to unwind.
5598 		 */
5599 		physical_rs->rs_start = logical_rs->rs_start;
5600 		physical_rs->rs_end = logical_rs->rs_end;
5601 
5602 		remain_rs->rs_start = logical_rs->rs_start;
5603 		remain_rs->rs_end = logical_rs->rs_start;
5604 
5605 		return;
5606 	}
5607 
5608 	vdev_t *pvd = vd->vdev_parent;
5609 	ASSERT3P(pvd, !=, NULL);
5610 	ASSERT3P(pvd->vdev_ops->vdev_op_xlate, !=, NULL);
5611 
5612 	/*
5613 	 * As this recursive function unwinds, translate the logical
5614 	 * range into its physical and any remaining components by calling
5615 	 * the vdev specific translate function.
5616 	 */
5617 	range_seg64_t intermediate = { 0 };
5618 	pvd->vdev_ops->vdev_op_xlate(vd, physical_rs, &intermediate, remain_rs);
5619 
5620 	physical_rs->rs_start = intermediate.rs_start;
5621 	physical_rs->rs_end = intermediate.rs_end;
5622 }
5623 
5624 void
5625 vdev_xlate_walk(vdev_t *vd, const range_seg64_t *logical_rs,
5626     vdev_xlate_func_t *func, void *arg)
5627 {
5628 	range_seg64_t iter_rs = *logical_rs;
5629 	range_seg64_t physical_rs;
5630 	range_seg64_t remain_rs;
5631 
5632 	while (!vdev_xlate_is_empty(&iter_rs)) {
5633 
5634 		vdev_xlate(vd, &iter_rs, &physical_rs, &remain_rs);
5635 
5636 		/*
5637 		 * With raidz and dRAID, it's possible that the logical range
5638 		 * does not live on this leaf vdev. Only when there is a non-
5639 		 * zero physical size call the provided function.
5640 		 */
5641 		if (!vdev_xlate_is_empty(&physical_rs))
5642 			func(arg, &physical_rs);
5643 
5644 		iter_rs = remain_rs;
5645 	}
5646 }
5647 
5648 static char *
5649 vdev_name(vdev_t *vd, char *buf, int buflen)
5650 {
5651 	if (vd->vdev_path == NULL) {
5652 		if (strcmp(vd->vdev_ops->vdev_op_type, "root") == 0) {
5653 			strlcpy(buf, vd->vdev_spa->spa_name, buflen);
5654 		} else if (!vd->vdev_ops->vdev_op_leaf) {
5655 			snprintf(buf, buflen, "%s-%llu",
5656 			    vd->vdev_ops->vdev_op_type,
5657 			    (u_longlong_t)vd->vdev_id);
5658 		}
5659 	} else {
5660 		strlcpy(buf, vd->vdev_path, buflen);
5661 	}
5662 	return (buf);
5663 }
5664 
5665 /*
5666  * Look at the vdev tree and determine whether any devices are currently being
5667  * replaced.
5668  */
5669 boolean_t
5670 vdev_replace_in_progress(vdev_t *vdev)
5671 {
5672 	ASSERT(spa_config_held(vdev->vdev_spa, SCL_ALL, RW_READER) != 0);
5673 
5674 	if (vdev->vdev_ops == &vdev_replacing_ops)
5675 		return (B_TRUE);
5676 
5677 	/*
5678 	 * A 'spare' vdev indicates that we have a replace in progress, unless
5679 	 * it has exactly two children, and the second, the hot spare, has
5680 	 * finished being resilvered.
5681 	 */
5682 	if (vdev->vdev_ops == &vdev_spare_ops && (vdev->vdev_children > 2 ||
5683 	    !vdev_dtl_empty(vdev->vdev_child[1], DTL_MISSING)))
5684 		return (B_TRUE);
5685 
5686 	for (int i = 0; i < vdev->vdev_children; i++) {
5687 		if (vdev_replace_in_progress(vdev->vdev_child[i]))
5688 			return (B_TRUE);
5689 	}
5690 
5691 	return (B_FALSE);
5692 }
5693 
5694 /*
5695  * Add a (source=src, propname=propval) list to an nvlist.
5696  */
5697 static void
5698 vdev_prop_add_list(nvlist_t *nvl, const char *propname, const char *strval,
5699     uint64_t intval, zprop_source_t src)
5700 {
5701 	nvlist_t *propval;
5702 
5703 	propval = fnvlist_alloc();
5704 	fnvlist_add_uint64(propval, ZPROP_SOURCE, src);
5705 
5706 	if (strval != NULL)
5707 		fnvlist_add_string(propval, ZPROP_VALUE, strval);
5708 	else
5709 		fnvlist_add_uint64(propval, ZPROP_VALUE, intval);
5710 
5711 	fnvlist_add_nvlist(nvl, propname, propval);
5712 	nvlist_free(propval);
5713 }
5714 
5715 static void
5716 vdev_props_set_sync(void *arg, dmu_tx_t *tx)
5717 {
5718 	vdev_t *vd;
5719 	nvlist_t *nvp = arg;
5720 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
5721 	objset_t *mos = spa->spa_meta_objset;
5722 	nvpair_t *elem = NULL;
5723 	uint64_t vdev_guid;
5724 	uint64_t objid;
5725 	nvlist_t *nvprops;
5726 
5727 	vdev_guid = fnvlist_lookup_uint64(nvp, ZPOOL_VDEV_PROPS_SET_VDEV);
5728 	nvprops = fnvlist_lookup_nvlist(nvp, ZPOOL_VDEV_PROPS_SET_PROPS);
5729 	vd = spa_lookup_by_guid(spa, vdev_guid, B_TRUE);
5730 
5731 	/* this vdev could get removed while waiting for this sync task */
5732 	if (vd == NULL)
5733 		return;
5734 
5735 	/*
5736 	 * Set vdev property values in the vdev props mos object.
5737 	 */
5738 	if (vd->vdev_root_zap != 0) {
5739 		objid = vd->vdev_root_zap;
5740 	} else if (vd->vdev_top_zap != 0) {
5741 		objid = vd->vdev_top_zap;
5742 	} else if (vd->vdev_leaf_zap != 0) {
5743 		objid = vd->vdev_leaf_zap;
5744 	} else {
5745 		panic("unexpected vdev type");
5746 	}
5747 
5748 	mutex_enter(&spa->spa_props_lock);
5749 
5750 	while ((elem = nvlist_next_nvpair(nvprops, elem)) != NULL) {
5751 		uint64_t intval;
5752 		const char *strval;
5753 		vdev_prop_t prop;
5754 		const char *propname = nvpair_name(elem);
5755 		zprop_type_t proptype;
5756 
5757 		switch (prop = vdev_name_to_prop(propname)) {
5758 		case VDEV_PROP_USERPROP:
5759 			if (vdev_prop_user(propname)) {
5760 				strval = fnvpair_value_string(elem);
5761 				if (strlen(strval) == 0) {
5762 					/* remove the property if value == "" */
5763 					(void) zap_remove(mos, objid, propname,
5764 					    tx);
5765 				} else {
5766 					VERIFY0(zap_update(mos, objid, propname,
5767 					    1, strlen(strval) + 1, strval, tx));
5768 				}
5769 				spa_history_log_internal(spa, "vdev set", tx,
5770 				    "vdev_guid=%llu: %s=%s",
5771 				    (u_longlong_t)vdev_guid, nvpair_name(elem),
5772 				    strval);
5773 			}
5774 			break;
5775 		default:
5776 			/* normalize the property name */
5777 			propname = vdev_prop_to_name(prop);
5778 			proptype = vdev_prop_get_type(prop);
5779 
5780 			if (nvpair_type(elem) == DATA_TYPE_STRING) {
5781 				ASSERT(proptype == PROP_TYPE_STRING);
5782 				strval = fnvpair_value_string(elem);
5783 				VERIFY0(zap_update(mos, objid, propname,
5784 				    1, strlen(strval) + 1, strval, tx));
5785 				spa_history_log_internal(spa, "vdev set", tx,
5786 				    "vdev_guid=%llu: %s=%s",
5787 				    (u_longlong_t)vdev_guid, nvpair_name(elem),
5788 				    strval);
5789 			} else if (nvpair_type(elem) == DATA_TYPE_UINT64) {
5790 				intval = fnvpair_value_uint64(elem);
5791 
5792 				if (proptype == PROP_TYPE_INDEX) {
5793 					const char *unused;
5794 					VERIFY0(vdev_prop_index_to_string(
5795 					    prop, intval, &unused));
5796 				}
5797 				VERIFY0(zap_update(mos, objid, propname,
5798 				    sizeof (uint64_t), 1, &intval, tx));
5799 				spa_history_log_internal(spa, "vdev set", tx,
5800 				    "vdev_guid=%llu: %s=%lld",
5801 				    (u_longlong_t)vdev_guid,
5802 				    nvpair_name(elem), (longlong_t)intval);
5803 			} else {
5804 				panic("invalid vdev property type %u",
5805 				    nvpair_type(elem));
5806 			}
5807 		}
5808 
5809 	}
5810 
5811 	mutex_exit(&spa->spa_props_lock);
5812 }
5813 
5814 int
5815 vdev_prop_set(vdev_t *vd, nvlist_t *innvl, nvlist_t *outnvl)
5816 {
5817 	spa_t *spa = vd->vdev_spa;
5818 	nvpair_t *elem = NULL;
5819 	uint64_t vdev_guid;
5820 	nvlist_t *nvprops;
5821 	int error = 0;
5822 
5823 	ASSERT(vd != NULL);
5824 
5825 	/* Check that vdev has a zap we can use */
5826 	if (vd->vdev_root_zap == 0 &&
5827 	    vd->vdev_top_zap == 0 &&
5828 	    vd->vdev_leaf_zap == 0)
5829 		return (SET_ERROR(EINVAL));
5830 
5831 	if (nvlist_lookup_uint64(innvl, ZPOOL_VDEV_PROPS_SET_VDEV,
5832 	    &vdev_guid) != 0)
5833 		return (SET_ERROR(EINVAL));
5834 
5835 	if (nvlist_lookup_nvlist(innvl, ZPOOL_VDEV_PROPS_SET_PROPS,
5836 	    &nvprops) != 0)
5837 		return (SET_ERROR(EINVAL));
5838 
5839 	if ((vd = spa_lookup_by_guid(spa, vdev_guid, B_TRUE)) == NULL)
5840 		return (SET_ERROR(EINVAL));
5841 
5842 	while ((elem = nvlist_next_nvpair(nvprops, elem)) != NULL) {
5843 		const char *propname = nvpair_name(elem);
5844 		vdev_prop_t prop = vdev_name_to_prop(propname);
5845 		uint64_t intval = 0;
5846 		const char *strval = NULL;
5847 
5848 		if (prop == VDEV_PROP_USERPROP && !vdev_prop_user(propname)) {
5849 			error = EINVAL;
5850 			goto end;
5851 		}
5852 
5853 		if (vdev_prop_readonly(prop)) {
5854 			error = EROFS;
5855 			goto end;
5856 		}
5857 
5858 		/* Special Processing */
5859 		switch (prop) {
5860 		case VDEV_PROP_PATH:
5861 			if (vd->vdev_path == NULL) {
5862 				error = EROFS;
5863 				break;
5864 			}
5865 			if (nvpair_value_string(elem, &strval) != 0) {
5866 				error = EINVAL;
5867 				break;
5868 			}
5869 			/* New path must start with /dev/ */
5870 			if (strncmp(strval, "/dev/", 5)) {
5871 				error = EINVAL;
5872 				break;
5873 			}
5874 			error = spa_vdev_setpath(spa, vdev_guid, strval);
5875 			break;
5876 		case VDEV_PROP_ALLOCATING:
5877 			if (nvpair_value_uint64(elem, &intval) != 0) {
5878 				error = EINVAL;
5879 				break;
5880 			}
5881 			if (intval != vd->vdev_noalloc)
5882 				break;
5883 			if (intval == 0)
5884 				error = spa_vdev_noalloc(spa, vdev_guid);
5885 			else
5886 				error = spa_vdev_alloc(spa, vdev_guid);
5887 			break;
5888 		case VDEV_PROP_FAILFAST:
5889 			if (nvpair_value_uint64(elem, &intval) != 0) {
5890 				error = EINVAL;
5891 				break;
5892 			}
5893 			vd->vdev_failfast = intval & 1;
5894 			break;
5895 		case VDEV_PROP_CHECKSUM_N:
5896 			if (nvpair_value_uint64(elem, &intval) != 0) {
5897 				error = EINVAL;
5898 				break;
5899 			}
5900 			vd->vdev_checksum_n = intval;
5901 			break;
5902 		case VDEV_PROP_CHECKSUM_T:
5903 			if (nvpair_value_uint64(elem, &intval) != 0) {
5904 				error = EINVAL;
5905 				break;
5906 			}
5907 			vd->vdev_checksum_t = intval;
5908 			break;
5909 		case VDEV_PROP_IO_N:
5910 			if (nvpair_value_uint64(elem, &intval) != 0) {
5911 				error = EINVAL;
5912 				break;
5913 			}
5914 			vd->vdev_io_n = intval;
5915 			break;
5916 		case VDEV_PROP_IO_T:
5917 			if (nvpair_value_uint64(elem, &intval) != 0) {
5918 				error = EINVAL;
5919 				break;
5920 			}
5921 			vd->vdev_io_t = intval;
5922 			break;
5923 		default:
5924 			/* Most processing is done in vdev_props_set_sync */
5925 			break;
5926 		}
5927 end:
5928 		if (error != 0) {
5929 			intval = error;
5930 			vdev_prop_add_list(outnvl, propname, strval, intval, 0);
5931 			return (error);
5932 		}
5933 	}
5934 
5935 	return (dsl_sync_task(spa->spa_name, NULL, vdev_props_set_sync,
5936 	    innvl, 6, ZFS_SPACE_CHECK_EXTRA_RESERVED));
5937 }
5938 
5939 int
5940 vdev_prop_get(vdev_t *vd, nvlist_t *innvl, nvlist_t *outnvl)
5941 {
5942 	spa_t *spa = vd->vdev_spa;
5943 	objset_t *mos = spa->spa_meta_objset;
5944 	int err = 0;
5945 	uint64_t objid;
5946 	uint64_t vdev_guid;
5947 	nvpair_t *elem = NULL;
5948 	nvlist_t *nvprops = NULL;
5949 	uint64_t intval = 0;
5950 	char *strval = NULL;
5951 	const char *propname = NULL;
5952 	vdev_prop_t prop;
5953 
5954 	ASSERT(vd != NULL);
5955 	ASSERT(mos != NULL);
5956 
5957 	if (nvlist_lookup_uint64(innvl, ZPOOL_VDEV_PROPS_GET_VDEV,
5958 	    &vdev_guid) != 0)
5959 		return (SET_ERROR(EINVAL));
5960 
5961 	nvlist_lookup_nvlist(innvl, ZPOOL_VDEV_PROPS_GET_PROPS, &nvprops);
5962 
5963 	if (vd->vdev_root_zap != 0) {
5964 		objid = vd->vdev_root_zap;
5965 	} else if (vd->vdev_top_zap != 0) {
5966 		objid = vd->vdev_top_zap;
5967 	} else if (vd->vdev_leaf_zap != 0) {
5968 		objid = vd->vdev_leaf_zap;
5969 	} else {
5970 		return (SET_ERROR(EINVAL));
5971 	}
5972 	ASSERT(objid != 0);
5973 
5974 	mutex_enter(&spa->spa_props_lock);
5975 
5976 	if (nvprops != NULL) {
5977 		char namebuf[64] = { 0 };
5978 
5979 		while ((elem = nvlist_next_nvpair(nvprops, elem)) != NULL) {
5980 			intval = 0;
5981 			strval = NULL;
5982 			propname = nvpair_name(elem);
5983 			prop = vdev_name_to_prop(propname);
5984 			zprop_source_t src = ZPROP_SRC_DEFAULT;
5985 			uint64_t integer_size, num_integers;
5986 
5987 			switch (prop) {
5988 			/* Special Read-only Properties */
5989 			case VDEV_PROP_NAME:
5990 				strval = vdev_name(vd, namebuf,
5991 				    sizeof (namebuf));
5992 				if (strval == NULL)
5993 					continue;
5994 				vdev_prop_add_list(outnvl, propname, strval, 0,
5995 				    ZPROP_SRC_NONE);
5996 				continue;
5997 			case VDEV_PROP_CAPACITY:
5998 				/* percent used */
5999 				intval = (vd->vdev_stat.vs_dspace == 0) ? 0 :
6000 				    (vd->vdev_stat.vs_alloc * 100 /
6001 				    vd->vdev_stat.vs_dspace);
6002 				vdev_prop_add_list(outnvl, propname, NULL,
6003 				    intval, ZPROP_SRC_NONE);
6004 				continue;
6005 			case VDEV_PROP_STATE:
6006 				vdev_prop_add_list(outnvl, propname, NULL,
6007 				    vd->vdev_state, ZPROP_SRC_NONE);
6008 				continue;
6009 			case VDEV_PROP_GUID:
6010 				vdev_prop_add_list(outnvl, propname, NULL,
6011 				    vd->vdev_guid, ZPROP_SRC_NONE);
6012 				continue;
6013 			case VDEV_PROP_ASIZE:
6014 				vdev_prop_add_list(outnvl, propname, NULL,
6015 				    vd->vdev_asize, ZPROP_SRC_NONE);
6016 				continue;
6017 			case VDEV_PROP_PSIZE:
6018 				vdev_prop_add_list(outnvl, propname, NULL,
6019 				    vd->vdev_psize, ZPROP_SRC_NONE);
6020 				continue;
6021 			case VDEV_PROP_ASHIFT:
6022 				vdev_prop_add_list(outnvl, propname, NULL,
6023 				    vd->vdev_ashift, ZPROP_SRC_NONE);
6024 				continue;
6025 			case VDEV_PROP_SIZE:
6026 				vdev_prop_add_list(outnvl, propname, NULL,
6027 				    vd->vdev_stat.vs_dspace, ZPROP_SRC_NONE);
6028 				continue;
6029 			case VDEV_PROP_FREE:
6030 				vdev_prop_add_list(outnvl, propname, NULL,
6031 				    vd->vdev_stat.vs_dspace -
6032 				    vd->vdev_stat.vs_alloc, ZPROP_SRC_NONE);
6033 				continue;
6034 			case VDEV_PROP_ALLOCATED:
6035 				vdev_prop_add_list(outnvl, propname, NULL,
6036 				    vd->vdev_stat.vs_alloc, ZPROP_SRC_NONE);
6037 				continue;
6038 			case VDEV_PROP_EXPANDSZ:
6039 				vdev_prop_add_list(outnvl, propname, NULL,
6040 				    vd->vdev_stat.vs_esize, ZPROP_SRC_NONE);
6041 				continue;
6042 			case VDEV_PROP_FRAGMENTATION:
6043 				vdev_prop_add_list(outnvl, propname, NULL,
6044 				    vd->vdev_stat.vs_fragmentation,
6045 				    ZPROP_SRC_NONE);
6046 				continue;
6047 			case VDEV_PROP_PARITY:
6048 				vdev_prop_add_list(outnvl, propname, NULL,
6049 				    vdev_get_nparity(vd), ZPROP_SRC_NONE);
6050 				continue;
6051 			case VDEV_PROP_PATH:
6052 				if (vd->vdev_path == NULL)
6053 					continue;
6054 				vdev_prop_add_list(outnvl, propname,
6055 				    vd->vdev_path, 0, ZPROP_SRC_NONE);
6056 				continue;
6057 			case VDEV_PROP_DEVID:
6058 				if (vd->vdev_devid == NULL)
6059 					continue;
6060 				vdev_prop_add_list(outnvl, propname,
6061 				    vd->vdev_devid, 0, ZPROP_SRC_NONE);
6062 				continue;
6063 			case VDEV_PROP_PHYS_PATH:
6064 				if (vd->vdev_physpath == NULL)
6065 					continue;
6066 				vdev_prop_add_list(outnvl, propname,
6067 				    vd->vdev_physpath, 0, ZPROP_SRC_NONE);
6068 				continue;
6069 			case VDEV_PROP_ENC_PATH:
6070 				if (vd->vdev_enc_sysfs_path == NULL)
6071 					continue;
6072 				vdev_prop_add_list(outnvl, propname,
6073 				    vd->vdev_enc_sysfs_path, 0, ZPROP_SRC_NONE);
6074 				continue;
6075 			case VDEV_PROP_FRU:
6076 				if (vd->vdev_fru == NULL)
6077 					continue;
6078 				vdev_prop_add_list(outnvl, propname,
6079 				    vd->vdev_fru, 0, ZPROP_SRC_NONE);
6080 				continue;
6081 			case VDEV_PROP_PARENT:
6082 				if (vd->vdev_parent != NULL) {
6083 					strval = vdev_name(vd->vdev_parent,
6084 					    namebuf, sizeof (namebuf));
6085 					vdev_prop_add_list(outnvl, propname,
6086 					    strval, 0, ZPROP_SRC_NONE);
6087 				}
6088 				continue;
6089 			case VDEV_PROP_CHILDREN:
6090 				if (vd->vdev_children > 0)
6091 					strval = kmem_zalloc(ZAP_MAXVALUELEN,
6092 					    KM_SLEEP);
6093 				for (uint64_t i = 0; i < vd->vdev_children;
6094 				    i++) {
6095 					const char *vname;
6096 
6097 					vname = vdev_name(vd->vdev_child[i],
6098 					    namebuf, sizeof (namebuf));
6099 					if (vname == NULL)
6100 						vname = "(unknown)";
6101 					if (strlen(strval) > 0)
6102 						strlcat(strval, ",",
6103 						    ZAP_MAXVALUELEN);
6104 					strlcat(strval, vname, ZAP_MAXVALUELEN);
6105 				}
6106 				if (strval != NULL) {
6107 					vdev_prop_add_list(outnvl, propname,
6108 					    strval, 0, ZPROP_SRC_NONE);
6109 					kmem_free(strval, ZAP_MAXVALUELEN);
6110 				}
6111 				continue;
6112 			case VDEV_PROP_NUMCHILDREN:
6113 				vdev_prop_add_list(outnvl, propname, NULL,
6114 				    vd->vdev_children, ZPROP_SRC_NONE);
6115 				continue;
6116 			case VDEV_PROP_READ_ERRORS:
6117 				vdev_prop_add_list(outnvl, propname, NULL,
6118 				    vd->vdev_stat.vs_read_errors,
6119 				    ZPROP_SRC_NONE);
6120 				continue;
6121 			case VDEV_PROP_WRITE_ERRORS:
6122 				vdev_prop_add_list(outnvl, propname, NULL,
6123 				    vd->vdev_stat.vs_write_errors,
6124 				    ZPROP_SRC_NONE);
6125 				continue;
6126 			case VDEV_PROP_CHECKSUM_ERRORS:
6127 				vdev_prop_add_list(outnvl, propname, NULL,
6128 				    vd->vdev_stat.vs_checksum_errors,
6129 				    ZPROP_SRC_NONE);
6130 				continue;
6131 			case VDEV_PROP_INITIALIZE_ERRORS:
6132 				vdev_prop_add_list(outnvl, propname, NULL,
6133 				    vd->vdev_stat.vs_initialize_errors,
6134 				    ZPROP_SRC_NONE);
6135 				continue;
6136 			case VDEV_PROP_OPS_NULL:
6137 				vdev_prop_add_list(outnvl, propname, NULL,
6138 				    vd->vdev_stat.vs_ops[ZIO_TYPE_NULL],
6139 				    ZPROP_SRC_NONE);
6140 				continue;
6141 			case VDEV_PROP_OPS_READ:
6142 				vdev_prop_add_list(outnvl, propname, NULL,
6143 				    vd->vdev_stat.vs_ops[ZIO_TYPE_READ],
6144 				    ZPROP_SRC_NONE);
6145 				continue;
6146 			case VDEV_PROP_OPS_WRITE:
6147 				vdev_prop_add_list(outnvl, propname, NULL,
6148 				    vd->vdev_stat.vs_ops[ZIO_TYPE_WRITE],
6149 				    ZPROP_SRC_NONE);
6150 				continue;
6151 			case VDEV_PROP_OPS_FREE:
6152 				vdev_prop_add_list(outnvl, propname, NULL,
6153 				    vd->vdev_stat.vs_ops[ZIO_TYPE_FREE],
6154 				    ZPROP_SRC_NONE);
6155 				continue;
6156 			case VDEV_PROP_OPS_CLAIM:
6157 				vdev_prop_add_list(outnvl, propname, NULL,
6158 				    vd->vdev_stat.vs_ops[ZIO_TYPE_CLAIM],
6159 				    ZPROP_SRC_NONE);
6160 				continue;
6161 			case VDEV_PROP_OPS_TRIM:
6162 				/*
6163 				 * TRIM ops and bytes are reported to user
6164 				 * space as ZIO_TYPE_IOCTL.  This is done to
6165 				 * preserve the vdev_stat_t structure layout
6166 				 * for user space.
6167 				 */
6168 				vdev_prop_add_list(outnvl, propname, NULL,
6169 				    vd->vdev_stat.vs_ops[ZIO_TYPE_IOCTL],
6170 				    ZPROP_SRC_NONE);
6171 				continue;
6172 			case VDEV_PROP_BYTES_NULL:
6173 				vdev_prop_add_list(outnvl, propname, NULL,
6174 				    vd->vdev_stat.vs_bytes[ZIO_TYPE_NULL],
6175 				    ZPROP_SRC_NONE);
6176 				continue;
6177 			case VDEV_PROP_BYTES_READ:
6178 				vdev_prop_add_list(outnvl, propname, NULL,
6179 				    vd->vdev_stat.vs_bytes[ZIO_TYPE_READ],
6180 				    ZPROP_SRC_NONE);
6181 				continue;
6182 			case VDEV_PROP_BYTES_WRITE:
6183 				vdev_prop_add_list(outnvl, propname, NULL,
6184 				    vd->vdev_stat.vs_bytes[ZIO_TYPE_WRITE],
6185 				    ZPROP_SRC_NONE);
6186 				continue;
6187 			case VDEV_PROP_BYTES_FREE:
6188 				vdev_prop_add_list(outnvl, propname, NULL,
6189 				    vd->vdev_stat.vs_bytes[ZIO_TYPE_FREE],
6190 				    ZPROP_SRC_NONE);
6191 				continue;
6192 			case VDEV_PROP_BYTES_CLAIM:
6193 				vdev_prop_add_list(outnvl, propname, NULL,
6194 				    vd->vdev_stat.vs_bytes[ZIO_TYPE_CLAIM],
6195 				    ZPROP_SRC_NONE);
6196 				continue;
6197 			case VDEV_PROP_BYTES_TRIM:
6198 				/*
6199 				 * TRIM ops and bytes are reported to user
6200 				 * space as ZIO_TYPE_IOCTL.  This is done to
6201 				 * preserve the vdev_stat_t structure layout
6202 				 * for user space.
6203 				 */
6204 				vdev_prop_add_list(outnvl, propname, NULL,
6205 				    vd->vdev_stat.vs_bytes[ZIO_TYPE_IOCTL],
6206 				    ZPROP_SRC_NONE);
6207 				continue;
6208 			case VDEV_PROP_REMOVING:
6209 				vdev_prop_add_list(outnvl, propname, NULL,
6210 				    vd->vdev_removing, ZPROP_SRC_NONE);
6211 				continue;
6212 			/* Numeric Properites */
6213 			case VDEV_PROP_ALLOCATING:
6214 				/* Leaf vdevs cannot have this property */
6215 				if (vd->vdev_mg == NULL &&
6216 				    vd->vdev_top != NULL) {
6217 					src = ZPROP_SRC_NONE;
6218 					intval = ZPROP_BOOLEAN_NA;
6219 				} else {
6220 					err = vdev_prop_get_int(vd, prop,
6221 					    &intval);
6222 					if (err && err != ENOENT)
6223 						break;
6224 
6225 					if (intval ==
6226 					    vdev_prop_default_numeric(prop))
6227 						src = ZPROP_SRC_DEFAULT;
6228 					else
6229 						src = ZPROP_SRC_LOCAL;
6230 				}
6231 
6232 				vdev_prop_add_list(outnvl, propname, NULL,
6233 				    intval, src);
6234 				break;
6235 			case VDEV_PROP_FAILFAST:
6236 				src = ZPROP_SRC_LOCAL;
6237 				strval = NULL;
6238 
6239 				err = zap_lookup(mos, objid, nvpair_name(elem),
6240 				    sizeof (uint64_t), 1, &intval);
6241 				if (err == ENOENT) {
6242 					intval = vdev_prop_default_numeric(
6243 					    prop);
6244 					err = 0;
6245 				} else if (err) {
6246 					break;
6247 				}
6248 				if (intval == vdev_prop_default_numeric(prop))
6249 					src = ZPROP_SRC_DEFAULT;
6250 
6251 				vdev_prop_add_list(outnvl, propname, strval,
6252 				    intval, src);
6253 				break;
6254 			case VDEV_PROP_CHECKSUM_N:
6255 			case VDEV_PROP_CHECKSUM_T:
6256 			case VDEV_PROP_IO_N:
6257 			case VDEV_PROP_IO_T:
6258 				err = vdev_prop_get_int(vd, prop, &intval);
6259 				if (err && err != ENOENT)
6260 					break;
6261 
6262 				if (intval == vdev_prop_default_numeric(prop))
6263 					src = ZPROP_SRC_DEFAULT;
6264 				else
6265 					src = ZPROP_SRC_LOCAL;
6266 
6267 				vdev_prop_add_list(outnvl, propname, NULL,
6268 				    intval, src);
6269 				break;
6270 			/* Text Properties */
6271 			case VDEV_PROP_COMMENT:
6272 				/* Exists in the ZAP below */
6273 				/* FALLTHRU */
6274 			case VDEV_PROP_USERPROP:
6275 				/* User Properites */
6276 				src = ZPROP_SRC_LOCAL;
6277 
6278 				err = zap_length(mos, objid, nvpair_name(elem),
6279 				    &integer_size, &num_integers);
6280 				if (err)
6281 					break;
6282 
6283 				switch (integer_size) {
6284 				case 8:
6285 					/* User properties cannot be integers */
6286 					err = EINVAL;
6287 					break;
6288 				case 1:
6289 					/* string property */
6290 					strval = kmem_alloc(num_integers,
6291 					    KM_SLEEP);
6292 					err = zap_lookup(mos, objid,
6293 					    nvpair_name(elem), 1,
6294 					    num_integers, strval);
6295 					if (err) {
6296 						kmem_free(strval,
6297 						    num_integers);
6298 						break;
6299 					}
6300 					vdev_prop_add_list(outnvl, propname,
6301 					    strval, 0, src);
6302 					kmem_free(strval, num_integers);
6303 					break;
6304 				}
6305 				break;
6306 			default:
6307 				err = ENOENT;
6308 				break;
6309 			}
6310 			if (err)
6311 				break;
6312 		}
6313 	} else {
6314 		/*
6315 		 * Get all properties from the MOS vdev property object.
6316 		 */
6317 		zap_cursor_t zc;
6318 		zap_attribute_t za;
6319 		for (zap_cursor_init(&zc, mos, objid);
6320 		    (err = zap_cursor_retrieve(&zc, &za)) == 0;
6321 		    zap_cursor_advance(&zc)) {
6322 			intval = 0;
6323 			strval = NULL;
6324 			zprop_source_t src = ZPROP_SRC_DEFAULT;
6325 			propname = za.za_name;
6326 
6327 			switch (za.za_integer_length) {
6328 			case 8:
6329 				/* We do not allow integer user properties */
6330 				/* This is likely an internal value */
6331 				break;
6332 			case 1:
6333 				/* string property */
6334 				strval = kmem_alloc(za.za_num_integers,
6335 				    KM_SLEEP);
6336 				err = zap_lookup(mos, objid, za.za_name, 1,
6337 				    za.za_num_integers, strval);
6338 				if (err) {
6339 					kmem_free(strval, za.za_num_integers);
6340 					break;
6341 				}
6342 				vdev_prop_add_list(outnvl, propname, strval, 0,
6343 				    src);
6344 				kmem_free(strval, za.za_num_integers);
6345 				break;
6346 
6347 			default:
6348 				break;
6349 			}
6350 		}
6351 		zap_cursor_fini(&zc);
6352 	}
6353 
6354 	mutex_exit(&spa->spa_props_lock);
6355 	if (err && err != ENOENT) {
6356 		return (err);
6357 	}
6358 
6359 	return (0);
6360 }
6361 
6362 EXPORT_SYMBOL(vdev_fault);
6363 EXPORT_SYMBOL(vdev_degrade);
6364 EXPORT_SYMBOL(vdev_online);
6365 EXPORT_SYMBOL(vdev_offline);
6366 EXPORT_SYMBOL(vdev_clear);
6367 
6368 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, default_ms_count, UINT, ZMOD_RW,
6369 	"Target number of metaslabs per top-level vdev");
6370 
6371 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, default_ms_shift, UINT, ZMOD_RW,
6372 	"Default lower limit for metaslab size");
6373 
6374 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, max_ms_shift, UINT, ZMOD_RW,
6375 	"Default upper limit for metaslab size");
6376 
6377 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, min_ms_count, UINT, ZMOD_RW,
6378 	"Minimum number of metaslabs per top-level vdev");
6379 
6380 ZFS_MODULE_PARAM(zfs_vdev, zfs_vdev_, ms_count_limit, UINT, ZMOD_RW,
6381 	"Practical upper limit of total metaslabs per top-level vdev");
6382 
6383 ZFS_MODULE_PARAM(zfs, zfs_, slow_io_events_per_second, UINT, ZMOD_RW,
6384 	"Rate limit slow IO (delay) events to this many per second");
6385 
6386 /* BEGIN CSTYLED */
6387 ZFS_MODULE_PARAM(zfs, zfs_, checksum_events_per_second, UINT, ZMOD_RW,
6388 	"Rate limit checksum events to this many checksum errors per second "
6389 	"(do not set below ZED threshold).");
6390 /* END CSTYLED */
6391 
6392 ZFS_MODULE_PARAM(zfs, zfs_, scan_ignore_errors, INT, ZMOD_RW,
6393 	"Ignore errors during resilver/scrub");
6394 
6395 ZFS_MODULE_PARAM(zfs_vdev, vdev_, validate_skip, INT, ZMOD_RW,
6396 	"Bypass vdev_validate()");
6397 
6398 ZFS_MODULE_PARAM(zfs, zfs_, nocacheflush, INT, ZMOD_RW,
6399 	"Disable cache flushes");
6400 
6401 ZFS_MODULE_PARAM(zfs, zfs_, embedded_slog_min_ms, UINT, ZMOD_RW,
6402 	"Minimum number of metaslabs required to dedicate one for log blocks");
6403 
6404 /* BEGIN CSTYLED */
6405 ZFS_MODULE_PARAM_CALL(zfs_vdev, zfs_vdev_, min_auto_ashift,
6406 	param_set_min_auto_ashift, param_get_uint, ZMOD_RW,
6407 	"Minimum ashift used when creating new top-level vdevs");
6408 
6409 ZFS_MODULE_PARAM_CALL(zfs_vdev, zfs_vdev_, max_auto_ashift,
6410 	param_set_max_auto_ashift, param_get_uint, ZMOD_RW,
6411 	"Maximum ashift used when optimizing for logical -> physical sector "
6412 	"size on new top-level vdevs");
6413 /* END CSTYLED */
6414